From 87dfa07d1ecc48c2017bbb9785da91f371b25bcb Mon Sep 17 00:00:00 2001 From: Dundy Pasupuleti Date: Wed, 11 Nov 2026 10:00:06 -0500 Subject: [PATCH 1/4] feat(endpoint): per-URL headers for repeated --url targets Add --url-header INDEX:Header:Value (YAML: endpoint.url_headers, a list aligned with endpoint.urls) so each round-robin target can carry its own headers, e.g. X-data-parallel-rank to pin targets to vLLM DP ranks. Global --header values still apply to every target; a per-URL header replaces a global one with the same name (case-insensitive). Headers are picked by the request's url_index, so repeated identical URLs stay distinct targets and multi-turn sessions keep their headers. Indices past the URL list and credential-bearing header names are rejected, since per-URL headers bypass redaction and secret transport. Closes #1125 Signed-off-by: Dundy Pasupuleti --- .../templates/crd-aiperfjob.yaml | 11 + .../templates/crd-aiperfsweep.yaml | 11 + docs/cli-options.md | 16 ++ docs/tutorials/multi-url-load-balancing.md | 37 +++ .../common/models/model_endpoint_info.py | 26 +++ src/aiperf/config/endpoint.py | 45 +++- .../config/flags/_converter_endpoint.py | 1 + src/aiperf/config/flags/_section_fields.py | 1 + src/aiperf/config/flags/cli_config.py | 21 ++ src/aiperf/config/loader/parsing.py | 85 +++++++ .../config/schema/aiperf-config.schema.json | 2 +- src/aiperf/endpoints/anthropic_messages.py | 3 +- src/aiperf/endpoints/base_endpoint.py | 2 +- .../config/test_config_override_invariants.py | 1 + tests/unit/config/test_url_headers.py | 221 ++++++++++++++++++ .../endpoints/test_endpoint_url_headers.py | 100 ++++++++ tools/ergonomics_baseline.json | 5 + 17 files changed, 583 insertions(+), 5 deletions(-) create mode 100644 tests/unit/config/test_url_headers.py create mode 100644 tests/unit/endpoints/test_endpoint_url_headers.py diff --git a/deploy/helm/aiperf-operator/templates/crd-aiperfjob.yaml b/deploy/helm/aiperf-operator/templates/crd-aiperfjob.yaml index 2a6332dc97..2e89032207 100644 --- a/deploy/helm/aiperf-operator/templates/crd-aiperfjob.yaml +++ b/deploy/helm/aiperf-operator/templates/crd-aiperfjob.yaml @@ -330,6 +330,17 @@ spec: routing. Values support environment variable substitution. additionalProperties: type: string + urlHeaders: + type: array + description: 'Extra HTTP headers for individual targets: entry ``i`` applies to requests sent to ``urls[i]`` + and may be ``{}``. Merged over ``headers``; on a name clash (case-insensitive) the per-URL value + wins. Use for routing hints such as ``X-data-parallel-rank``. Credential-bearing header names are + rejected; put credentials in ``headers`` or ``api_key``. Example: [{X-data-parallel-rank: ''0''}, + {X-data-parallel-rank: ''1''}]' + items: + type: object + additionalProperties: + type: string extra: type: object description: 'Additional fields to include in request body. Merged into every request. Common fields: diff --git a/deploy/helm/aiperf-operator/templates/crd-aiperfsweep.yaml b/deploy/helm/aiperf-operator/templates/crd-aiperfsweep.yaml index 4689735662..412d1b649e 100644 --- a/deploy/helm/aiperf-operator/templates/crd-aiperfsweep.yaml +++ b/deploy/helm/aiperf-operator/templates/crd-aiperfsweep.yaml @@ -469,6 +469,17 @@ spec: routing. Values support environment variable substitution. additionalProperties: type: string + urlHeaders: + type: array + description: 'Extra HTTP headers for individual targets: entry ``i`` applies to requests sent to ``urls[i]`` + and may be ``{}``. Merged over ``headers``; on a name clash (case-insensitive) the per-URL value + wins. Use for routing hints such as ``X-data-parallel-rank``. Credential-bearing header names are + rejected; put credentials in ``headers`` or ``api_key``. Example: [{X-data-parallel-rank: ''0''}, + {X-data-parallel-rank: ''1''}]' + items: + type: object + additionalProperties: + type: string extra: type: object description: 'Additional fields to include in request body. Merged into every request. Common fields: diff --git a/docs/cli-options.md b/docs/cli-options.md index c0250820f1..c1342960b4 100644 --- a/docs/cli-options.md +++ b/docs/cli-options.md @@ -506,6 +506,10 @@ Strategy for selecting URLs when multiple `--url` values are provided. 'round_ro
_Choices: [`round_robin`]_
_Default: `round_robin`_ +#### `--url-header` `` + +HTTP header to send only to one `--url` target, as `INDEX:Header:Value`, where `INDEX` is the 0-based position of the URL in the `--url` list. Can be specified multiple times. Global `--header` values still apply to every target; on a name clash (case-insensitive) the per-URL value wins. Repeated identical URLs stay distinct targets, so `--url http://localhost:8000 --url http://localhost:8000 --url-header 0:X-data-parallel-rank:0 --url-header 1:X-data-parallel-rank:1` pins each round-robin target to a vLLM data-parallel rank. Values are sent as given (never coerced). Credential-bearing header names are rejected here; pass credentials with `--header` or `--api-key`. + #### `--request-timeout-seconds` `` Maximum time in seconds to wait for each HTTP request to complete, including connection establishment, request transmission, and response receipt. Applies to both streaming and non-streaming requests. Requests exceeding this timeout are cancelled and recorded as failures. @@ -2176,6 +2180,10 @@ Strategy for selecting URLs when multiple `--url` values are provided. 'round_ro
_Choices: [`round_robin`]_
_Default: `round_robin`_ +#### `--url-header` `` + +HTTP header to send only to one `--url` target, as `INDEX:Header:Value`, where `INDEX` is the 0-based position of the URL in the `--url` list. Can be specified multiple times. Global `--header` values still apply to every target; on a name clash (case-insensitive) the per-URL value wins. Repeated identical URLs stay distinct targets, so `--url http://localhost:8000 --url http://localhost:8000 --url-header 0:X-data-parallel-rank:0 --url-header 1:X-data-parallel-rank:1` pins each round-robin target to a vLLM data-parallel rank. Values are sent as given (never coerced). Credential-bearing header names are rejected here; pass credentials with `--header` or `--api-key`. + #### `--request-timeout-seconds` `` Maximum time in seconds to wait for each HTTP request to complete, including connection establishment, request transmission, and response receipt. Applies to both streaming and non-streaming requests. Requests exceeding this timeout are cancelled and recorded as failures. @@ -3841,6 +3849,10 @@ Strategy for selecting URLs when multiple `--url` values are provided. 'round_ro
_Choices: [`round_robin`]_
_Default: `round_robin`_ +#### `--url-header` `` + +HTTP header to send only to one `--url` target, as `INDEX:Header:Value`, where `INDEX` is the 0-based position of the URL in the `--url` list. Can be specified multiple times. Global `--header` values still apply to every target; on a name clash (case-insensitive) the per-URL value wins. Repeated identical URLs stay distinct targets, so `--url http://localhost:8000 --url http://localhost:8000 --url-header 0:X-data-parallel-rank:0 --url-header 1:X-data-parallel-rank:1` pins each round-robin target to a vLLM data-parallel rank. Values are sent as given (never coerced). Credential-bearing header names are rejected here; pass credentials with `--header` or `--api-key`. + #### `--request-timeout-seconds` `` Maximum time in seconds to wait for each HTTP request to complete, including connection establishment, request transmission, and response receipt. Applies to both streaming and non-streaming requests. Requests exceeding this timeout are cancelled and recorded as failures. @@ -6013,6 +6025,10 @@ Strategy for selecting URLs when multiple `--url` values are provided. 'round_ro
_Choices: [`round_robin`]_
_Default: `round_robin`_ +#### `--url-header` `` + +HTTP header to send only to one `--url` target, as `INDEX:Header:Value`, where `INDEX` is the 0-based position of the URL in the `--url` list. Can be specified multiple times. Global `--header` values still apply to every target; on a name clash (case-insensitive) the per-URL value wins. Repeated identical URLs stay distinct targets, so `--url http://localhost:8000 --url http://localhost:8000 --url-header 0:X-data-parallel-rank:0 --url-header 1:X-data-parallel-rank:1` pins each round-robin target to a vLLM data-parallel rank. Values are sent as given (never coerced). Credential-bearing header names are rejected here; pass credentials with `--header` or `--api-key`. + #### `--request-timeout-seconds` `` Maximum time in seconds to wait for each HTTP request to complete, including connection establishment, request transmission, and response receipt. Applies to both streaming and non-streaming requests. Requests exceeding this timeout are cancelled and recorded as failures. diff --git a/docs/tutorials/multi-url-load-balancing.md b/docs/tutorials/multi-url-load-balancing.md index 0c56af3eee..518d916cae 100644 --- a/docs/tutorials/multi-url-load-balancing.md +++ b/docs/tutorials/multi-url-load-balancing.md @@ -100,12 +100,49 @@ aiperf profile --model llama \ --request-count 100 ``` +## Per-URL Headers + +`--header` sends the same headers to every URL. To send a header to only one target, use `--url-header INDEX:Header:Value`, where `INDEX` is the 0-based position of the URL in the `--url` list. Repeated identical URLs stay distinct round-robin targets, so this pins each target to a vLLM data-parallel rank through vLLM's `X-data-parallel-rank` header, without the DP load balancer in the measurement: + +```bash +# Server started with: vllm serve llama --data-parallel-size 2 +aiperf profile --model llama \ + --url http://localhost:8000 \ + --url http://localhost:8000 \ + --url-header 0:X-data-parallel-rank:0 \ + --url-header 1:X-data-parallel-rank:1 \ + --concurrency 16 \ + --request-count 200 +``` + +In a config file, `endpoint.url_headers` is a list aligned with `endpoint.urls`; use `{}` for a URL without extra headers: + +```yaml +benchmark: + endpoint: + urls: [http://localhost:8000, http://localhost:8000, http://localhost:8000] + headers: + X-Team: perf + url_headers: + - {X-data-parallel-rank: "0"} + - {} + - {X-data-parallel-rank: "2"} +``` + +- Global headers still apply to every URL. A per-URL header replaces a global header with the same name (compared case-insensitively). +- A session keeps its URL for every turn, so all turns of a multi-turn conversation carry the same per-URL headers. +- Values are sent as given; unlike `--header`, they are never coerced to numbers or booleans. +- An index past the last `--url` is a configuration error. +- Credential-bearing header names (for example `Authorization` or names containing `token` or `api-key`) are rejected: per-URL headers are not redacted from artifacts or transported as secrets. Pass credentials with `--header` or `--api-key`. +- Per-URL headers apply to inference requests. Readiness probes and control hooks (`--reset-kv-cache`, `--server-profiler`) keep using the global headers. + ## CLI Options | Option | Type | Default | Description | |--------|------|---------|-------------| | `--url` | list | localhost:8000 | One or more endpoint URLs; multiple URLs enable load balancing | | `--url-strategy` | enum | round_robin | Strategy for distributing requests across multiple URLs | +| `--url-header` | list | none | `INDEX:Header:Value` header sent only to the URL at that 0-based position; repeatable | ## Behavior Notes diff --git a/src/aiperf/common/models/model_endpoint_info.py b/src/aiperf/common/models/model_endpoint_info.py index cd9ce08244..cbef2138d6 100644 --- a/src/aiperf/common/models/model_endpoint_info.py +++ b/src/aiperf/common/models/model_endpoint_info.py @@ -99,6 +99,12 @@ def _redact_headers(self, value: list[tuple[str, str]]) -> list[tuple[str, str]] return redact_header_tuples(value) + url_headers: list[dict[str, str]] = Field( + default=[], + description="Extra headers for individual URLs: entry i applies to " + "base_urls[i]. Merged over headers for requests sent to that URL.", + ) + api_key: str | None = Field( default=EndpointDefaults.API_KEY, description="API key to use for the endpoint.", @@ -218,6 +224,25 @@ def get_url(self, index: int | None = None) -> str: return self.base_urls[0] return self.base_urls[index % len(self.base_urls)] + def get_headers(self, index: int | None = None) -> dict[str, str]: + """Get the custom headers for the URL at ``index``. + + Global ``headers`` apply to every URL; the URL's ``url_headers`` are laid + over them, replacing a global header of the same name regardless of + casing so the request never carries both. ``index`` wraps around like + :meth:`get_url`. + """ + headers = dict(self.headers or ()) + position = 0 if index is None else index % len(self.base_urls) + overrides = ( + self.url_headers[position] if position < len(self.url_headers) else None + ) + if overrides: + replaced = {name.lower() for name in overrides} + headers = {k: v for k, v in headers.items() if k.lower() not in replaced} + headers.update(overrides) + return headers + class ModelEndpointInfo(AIPerfBaseModel): """Information about a model endpoint.""" @@ -264,6 +289,7 @@ def from_config(cls, cfg: BenchmarkConfig) -> ModelEndpointInfo: allow_empty_content=ep.allow_empty_content, base_urls=list(ep.urls), headers=list((getattr(ep, "headers", {}) or {}).items()), + url_headers=list(getattr(ep, "url_headers", []) or []), extra=list((getattr(ep, "extra", {}) or {}).items()), timeout=ep.timeout, api_key=ep.api_key, diff --git a/src/aiperf/config/endpoint.py b/src/aiperf/config/endpoint.py index 87aa7fd8dd..814990877c 100644 --- a/src/aiperf/config/endpoint.py +++ b/src/aiperf/config/endpoint.py @@ -16,6 +16,7 @@ from pydantic import ( AfterValidator, + BeforeValidator, ConfigDict, Field, PrivateAttr, @@ -37,7 +38,7 @@ parse_enabled_or_config, require_relative_path, ) -from aiperf.config.loader.parsing import normalize_http_urls +from aiperf.config.loader.parsing import normalize_http_urls, parse_url_headers from aiperf.config.sagemaker import SageMakerConfig from aiperf.plugin.enums import ( EndpointType, @@ -546,6 +547,21 @@ def _redact_headers(self, value: dict[str, str]) -> dict[str, str]: return redact_headers(value) or {} + url_headers: Annotated[ + list[dict[str, str]], + Field( + default_factory=list, + description="Extra HTTP headers for individual targets: entry ``i`` " + "applies to requests sent to ``urls[i]`` and may be ``{}``. Merged " + "over ``headers``; on a name clash (case-insensitive) the per-URL " + "value wins. Use for routing hints such as ``X-data-parallel-rank``. " + "Credential-bearing header names are rejected; put credentials in " + "``headers`` or ``api_key``. " + "Example: [{X-data-parallel-rank: '0'}, {X-data-parallel-rank: '1'}]", + ), + BeforeValidator(parse_url_headers), + ] + extra: Annotated[ dict[str, Any], Field( @@ -831,6 +847,33 @@ def _validate_endpoint_boundaries(self) -> Self: raise ValueError("endpoint.path must start with a leading slash") return self + @model_validator(mode="after") + def _validate_url_headers(self) -> Self: + """Bind per-URL headers to existing URLs and keep credentials out of them. + + Credential-bearing headers are redacted from serialized configs and + rehydrated out of band only for ``headers`` (``AIPERF_INJECTED_HEADERS``), + so a secret here would reach worker processes as ````. + """ + from aiperf.common.redact import is_sensitive_header_name + + if len(self.url_headers) > len(self.urls): + raise ValueError( + f"url_headers: URL index {len(self.url_headers) - 1} is out of " + f"range; {len(self.urls)} URL(s) are configured, so valid " + f"indices are 0..{len(self.urls) - 1}." + ) + for index, headers in enumerate(self.url_headers): + sensitive = sorted(n for n in headers if is_sensitive_header_name(n)) + if sensitive: + raise ValueError( + f"url_headers: {', '.join(sensitive)} for URL {index} looks " + "credential-bearing. Per-URL headers are not redacted or " + "transported as secrets; pass credentials with --header or " + "--api-key instead." + ) + return self + @model_validator(mode="after") def _validate_template_required(self) -> Self: if self.type == EndpointType.TEMPLATE and self.template is None: diff --git a/src/aiperf/config/flags/_converter_endpoint.py b/src/aiperf/config/flags/_converter_endpoint.py index 934fd3dc2f..5060baa103 100644 --- a/src/aiperf/config/flags/_converter_endpoint.py +++ b/src/aiperf/config/flags/_converter_endpoint.py @@ -59,6 +59,7 @@ def _endpoint_template_fallback(endpoint: dict[str, Any]) -> None: # Map (CLIConfig endpoint field name) -> (AIPerfConfig endpoint key). _ENDPOINT_FIELD_MAP: dict[str, str] = { "url_selection_strategy": "url_strategy", + "url_headers": "url_headers", "endpoint_type": "type", "streaming": "streaming", "allow_empty_content": "allow_empty_content", diff --git a/src/aiperf/config/flags/_section_fields.py b/src/aiperf/config/flags/_section_fields.py index 8ff3a9b4a8..fe9c50683f 100644 --- a/src/aiperf/config/flags/_section_fields.py +++ b/src/aiperf/config/flags/_section_fields.py @@ -47,6 +47,7 @@ "transport", "endpoint_type", "url_selection_strategy", + "url_headers", "per_chunk_usage", "urls", "use_legacy_max_tokens", diff --git a/src/aiperf/config/flags/cli_config.py b/src/aiperf/config/flags/cli_config.py index ebd07b20ef..5d6da31658 100644 --- a/src/aiperf/config/flags/cli_config.py +++ b/src/aiperf/config/flags/cli_config.py @@ -71,6 +71,7 @@ parse_str_or_dict_as_tuple_list, parse_str_or_list, parse_str_or_list_of_positive_values, + parse_url_headers, raise_rejected_input_file, reject_moved_custom_dataset_type, require_turn_mean_at_least_one, @@ -234,6 +235,26 @@ class CLIConfig(BaseConfig): ), ] = EndpointDefaults.URL_STRATEGY + url_headers: Annotated[ + Any, + Field( + description="HTTP header to send only to one `--url` target, as `INDEX:Header:Value`, where `INDEX` " + "is the 0-based position of the URL in the `--url` list. Can be specified multiple times. " + "Global `--header` values still apply to every target; on a name clash (case-insensitive) the " + "per-URL value wins. Repeated identical URLs stay distinct targets, so " + "`--url http://localhost:8000 --url http://localhost:8000 --url-header 0:X-data-parallel-rank:0 " + "--url-header 1:X-data-parallel-rank:1` pins each round-robin target to a vLLM data-parallel rank. " + "Values are sent as given (never coerced). Credential-bearing header names are rejected here; " + "pass credentials with `--header` or `--api-key`.", + ), + BeforeValidator(parse_url_headers), + CLIParameter( + name=("--url-header",), + consume_multiple=True, + group=Groups.ENDPOINT, + ), + ] = [] + timeout_seconds: Annotated[ float, Field( diff --git a/src/aiperf/config/loader/parsing.py b/src/aiperf/config/loader/parsing.py index f5186a895e..01a3624441 100644 --- a/src/aiperf/config/loader/parsing.py +++ b/src/aiperf/config/loader/parsing.py @@ -242,6 +242,91 @@ def parse_str_or_dict_as_tuple_list(input: Any | None) -> list[tuple[str, Any]] raise ValueError(f"User Config: {input} - must be a valid string, list, or dict") +def _parse_url_index(raw: Any, source: Any) -> int: + """Parse a per-URL header index (the 0-based position in the URL list).""" + if isinstance(raw, int) and not isinstance(raw, bool) and raw >= 0: + return raw + if isinstance(raw, str) and raw.strip().isdigit(): + return int(raw.strip()) + raise ValueError( + f"User Config: {source!r} - URL index must be a non-negative integer" + ) + + +def _set_url_headers( + output: list[dict[str, str]], index: int, headers: Any, source: Any +) -> None: + if not isinstance(headers, dict): + raise ValueError( + f"User Config: {source!r} - headers for URL {index} must be a " + "mapping of header name to value" + ) + while len(output) <= index: + output.append({}) + for name, value in headers.items(): + name = str(name).strip() + if not name: + raise ValueError( + f"User Config: {source!r} - header name for URL {index} " + "must not be empty" + ) + output[index][name] = str(value) + + +def _merge_url_headers_item(output: list[dict[str, str]], item: Any) -> None: + if isinstance(item, str) and item.lstrip().startswith(("{", "[")): + try: + item = load_json_str(item) + except orjson.JSONDecodeError as e: + raise ValueError( + f"User Config: {item} - must be a valid JSON string" + ) from e + if isinstance(item, list): + for index, headers in enumerate(item): + _set_url_headers(output, index, headers, item) + return + if isinstance(item, dict): + for raw_index, headers in item.items(): + _set_url_headers(output, _parse_url_index(raw_index, item), headers, item) + elif isinstance(item, str): + parts = item.split(":", 2) + if len(parts) != 3: + raise ValueError( + f"User Config: {item!r} - must be in 'INDEX:Header:Value' " + "format, e.g. '0:X-data-parallel-rank:0'" + ) + index = _parse_url_index(parts[0], item) + _set_url_headers(output, index, {parts[1]: parts[2].strip()}, item) + else: + raise ValueError( + f"User Config: {item!r} - must be an 'INDEX:Header:Value' string, " + "a list of header mappings, or a mapping of URL index to headers" + ) + + +def parse_url_headers(input: Any) -> list[dict[str, str]] | None: + """Parse per-URL headers into a list aligned with the URL list. + + Entry ``i`` holds the headers for the ``i``-th URL; URLs without headers + get ``{}``. Accepts a list of header mappings (the YAML form), the + ``INDEX:Header:Value`` CLI form (the value may contain colons), a mapping + of URL index to headers, a JSON string of either, or a list of CLI + strings. Later entries for the same header win. Values are kept as + strings: unlike ``--header`` they are never coerced, so + ``0:X-data-parallel-rank:0`` sends ``"0"``. + """ + if input is None: + return None + output: list[dict[str, str]] = [] + if isinstance(input, (list, tuple)) and all(isinstance(i, dict) for i in input): + for index, headers in enumerate(input): + _set_url_headers(output, index, headers, input) + return output + for item in input if isinstance(input, (list, tuple)) else [input]: + _merge_url_headers_item(output, item) + return output + + def print_str_or_list(input: Any) -> str: """Convert a list, Enum, or scalar to a display string. diff --git a/src/aiperf/config/schema/aiperf-config.schema.json b/src/aiperf/config/schema/aiperf-config.schema.json index b87aef746d..8c1d196c93 100644 --- a/src/aiperf/config/schema/aiperf-config.schema.json +++ b/src/aiperf/config/schema/aiperf-config.schema.json @@ -1 +1 @@ -{"$schema":"https://json-schema.org/draft/2020-12/schema","$id":"https://nvidia.github.io/aiperf/schemas/aiperf-config.schema.json","$defs":{"AIPerfLogLevel":{"description":"Logging levels for AIPerf output verbosity.","enum":["TRACE","DEBUG","INFO","NOTICE","WARNING","SUCCESS","ERROR","CRITICAL"],"title":"AIPerfLogLevel","type":"string"},"AccuracyBenchmarkType":{"enum":["mmlu","mmlu_pro","aime","hellaswag","bigbench","aime24","aime25","math_500","gsm8k","gpqa_diamond","lcb_codegeneration"],"title":"AccuracyBenchmarkType","type":"string"},"AccuracyConfig":{"additionalProperties":false,"description":"Configuration for accuracy benchmarking mode.\n\nWhen benchmark is set, enables accuracy evaluation alongside\nperformance profiling using standard benchmarks (MMLU, AIME, etc.).","properties":{"benchmark":{"anyOf":[{"$ref":"#/$defs/AccuracyBenchmarkType"},{"type":"null"}],"default":null,"description":"Accuracy benchmark to run. When set, enables accuracy benchmarking alongside performance profiling. AIME variants: 'aime' is the legacy combined set (deprecated for new runs); prefer the year-pinned 'aime24' or 'aime25' for reproducibility."},"tasks":{"anyOf":[{"type":"string"},{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Specific tasks or subtasks within the benchmark to evaluate (e.g., specific MMLU subjects). Accepts comma-separated values (e.g. abstract_algebra,anatomy) or repeated flags. If not set, all tasks are included.","title":"Tasks"},"nShots":{"anyOf":[{"maximum":32,"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of few-shot examples to include in the prompt. 0 means zero-shot evaluation, None uses the benchmark default (e.g. MMLU=5). Maximum 32.","title":"Nshots","x-jinja2-supported":true},"enableCot":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Enable chain-of-thought prompting for accuracy evaluation. Adds reasoning instructions to the prompt. Defaults to the benchmark's ``default_enable_cot`` metadata when unset (e.g. AIME defaults to True).","title":"Enablecot"},"grader":{"anyOf":[{"$ref":"#/$defs/AccuracyGraderType"},{"type":"null"}],"default":null,"description":"Override the default grader for the selected benchmark (e.g., exact_match, math, multiple_choice, code_execution). If not set, uses the benchmark's default grader."},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Custom system prompt to use for accuracy evaluation. Overrides any benchmark-specific system prompt.","title":"Systemprompt"},"verbose":{"default":false,"description":"Enable verbose output for accuracy evaluation, showing per-problem grading details.","title":"Verbose","type":"boolean"}},"title":"AccuracyConfig","type":"object"},"AccuracyGraderType":{"enum":["exact_match","math","multiple_choice","code_execution","lighteval_expr","lighteval_latex","lighteval_gpqa","lighteval_gsm8k","mmlu_pro"],"title":"AccuracyGraderType","type":"string"},"AdaptiveSearchSweep":{"additionalProperties":false,"description":"Adaptive outer-loop search (Bayesian / monotonic / etc.).\n\nA third sweep variant: BO drives variation choice itself, so the\ngrid-style ``iteration_order`` and ``same_seed`` knobs do not apply.\nInherits cooldown/sla_filters/post_process from `_SweepBase` only.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"type":{"const":"adaptive_search","default":"adaptive_search","description":"Sweep type discriminator.","title":"Type","type":"string"},"planner":{"$ref":"#/$defs/SearchPlannerType","default":"bayesian","description":"Outer-loop planner plugin name (bayesian | optuna | monotonic_sla | smooth_isotonic)."},"searchSpace":{"description":"Dimensions to optimize over.","items":{"$ref":"#/$defs/SearchSpaceDimension"},"minItems":1,"title":"Searchspace","type":"array"},"objectives":{"description":"What to optimize. Length-1 = single-objective search; length-N = multi-objective search. BoTorch-backed Pareto BO requires --optuna-sampler botorch with a multi-objective acquisition like qlognehvi.","items":{"$ref":"#/$defs/Objective"},"minItems":1,"title":"Objectives","type":"array"},"outcomeConstraints":{"description":"Feasibility gates on non-optimized metrics. Empty = no constraints.","items":{"$ref":"#/$defs/OutcomeConstraint"},"title":"Outcomeconstraints","type":"array"},"maxIterations":{"description":"Maximum BO iterations.","title":"Maxiterations","oneOf":[{"description":"Maximum BO iterations.","maximum":200,"minimum":2,"title":"Maxiterations","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"nInitialPoints":{"default":5,"description":"Sobol-random points before the GP fits. Must be < max_iterations for `bayesian`/`optuna` planners; ignored by 1-D SLA planners.","title":"Ninitialpoints","oneOf":[{"default":5,"description":"Sobol-random points before the GP fits. Must be < max_iterations for `bayesian`/`optuna` planners; ignored by 1-D SLA planners.","minimum":1,"title":"Ninitialpoints","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"plateauWindow":{"default":8,"description":"Recent-iteration window for plateau detection. Default 8 keeps the window strictly larger than ``n_initial_points`` (default 5), so the plateau test always inspects at least a few GP-driven iterations rather than tripping on the Sobol-only prefix when the initial points happen to land in a flat region.","title":"Plateauwindow","oneOf":[{"default":8,"description":"Recent-iteration window for plateau detection. Default 8 keeps the window strictly larger than ``n_initial_points`` (default 5), so the plateau test always inspects at least a few GP-driven iterations rather than tripping on the Sobol-only prefix when the initial points happen to land in a flat region.","minimum":2,"title":"Plateauwindow","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"plateauThreshold":{"default":0.01,"description":"Coefficient-of-variation threshold for plateau stop.","title":"Plateauthreshold","oneOf":[{"default":0.01,"description":"Coefficient-of-variation threshold for plateau stop.","exclusiveMinimum":0,"title":"Plateauthreshold","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"improvementPatience":{"default":10,"description":"Stop after this many no-improvement iterations.","title":"Improvementpatience","oneOf":[{"default":10,"description":"Stop after this many no-improvement iterations.","minimum":2,"title":"Improvementpatience","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"randomSeed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"If set, passed as `random_state` to the planner backend.","title":"Randomseed","x-jinja2-supported":true},"optunaSampler":{"default":"botorch","description":"Optuna sampler when planner=optuna. Default ``botorch`` is the preferred Gaussian-process qLogNEI/qLogNEHVI path and requires the optional ``botorch`` extra; when this default is implicit and the optional stack is unavailable, the planner warns and falls back to ``tpe``. ``tpe`` is the dep-light Tree-Parzen estimator that ships with Optuna core. ``gp`` is Optuna's pure-GP sampler and requires torch.","enum":["gp","tpe","botorch"],"title":"Optunasampler","type":"string"},"optunaAcquisition":{"anyOf":[{"enum":["logei","qlogei","qnei","qlognei","qehvi","qnehvi","qlognehvi"],"type":"string"},{"type":"null"}],"default":null,"description":"Acquisition override for the BoTorch sampler. None lets Optuna pick (single-objective LogEI). ``qnei`` = Letham 2017 noisy-EI; ``qlognei`` = Ament 2023 (BoTorch's strongly recommended modern default; requires botorch>=0.10). ``qehvi``/``qnehvi``/``qlognehvi`` are multi-objective hypervolume acquisitions for Pareto BO. Only meaningful when planner=optuna and optuna_sampler=botorch; ignored otherwise.","title":"Optunaacquisition"},"optunaTerminator":{"default":"none","description":"Optional posterior-regret stopping rule (Makarova 2022 / Ishibashi 2023; same family as Wilson 2024). Layered on top of three-signal convergence. Only meaningful when planner=optuna; ignored otherwise.","enum":["regret","emmr","none"],"title":"Optunaterminator","type":"string"},"objectivePooling":{"default":"mean","description":"Per-trial percentile aggregation. ``mean`` = arithmetic mean of per-trial percentiles across trials (current default). ``pooled`` = walks each trial's profile_export.jsonl and computes np.percentile over the pooled raw-sample bag (the statistic that satisfies SLO claims; requires --export-level records). Only meaningful when objective.stat is a percentile (p50/p90/p95/p99); a no-op when stat=avg.","enum":["mean","pooled"],"title":"Objectivepooling","type":"string"},"monotonicStabilityTrials":{"default":2,"description":"MonotonicSLA planner stability-window size.","title":"Monotonicstabilitytrials","oneOf":[{"default":2,"description":"MonotonicSLA planner stability-window size.","minimum":1,"title":"Monotonicstabilitytrials","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaReplicates":{"default":0,"description":"Replicate-step trial count override for the smooth_isotonic SLA planner. 0 = auto (Hyperband-flavored formula based on sigma_margin/threshold). >0 = fixed override. Ignored by other planners.","title":"Slareplicates","oneOf":[{"default":0,"description":"Replicate-step trial count override for the smooth_isotonic SLA planner. 0 = auto (Hyperband-flavored formula based on sigma_margin/threshold). >0 = fixed override. Ignored by other planners.","minimum":0,"title":"Slareplicates","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaPrecision":{"default":"normal","description":"Per-probe sample-budget knob for the smooth_isotonic SLA planner. tight=10000, normal=1000, coarse=300 requests. Drives p99 CI width. Ignored by other planners.","enum":["tight","normal","coarse"],"title":"Slaprecision","type":"string"},"slaWarmupSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Discard this many seconds at the start of each probe before computing margins, to skip CUDA-graph warmup and KV-cache shape transients. None = auto (max(30s, 3 * inter-batch-time)). Ignored by other planners.","title":"Slawarmupseconds","x-jinja2-supported":true},"constraintMode":{"default":"penalty","description":"Constrained-BO acquisition mode (penalty | eic).","enum":["penalty","eic"],"title":"Constraintmode","type":"string"},"slaTiers":{"description":"Multi-tier SLO definitions parsed from --search-sla-tier flags. Each tier is a named group of SLA filters sharing a common boundary search. When non-empty (2-10 tiers), the MultiTierPlanner is activated instead of single-tier planners.","items":{"$ref":"#/$defs/SLOTier"},"title":"Slatiers","type":"array"}},"required":["searchSpace","objectives","maxIterations"],"title":"AdaptiveSearchSweep","type":"object"},"ArtifactsConfig":{"additionalProperties":false,"description":"Artifacts configuration for benchmark output.\n\nControls where and how benchmark results are exported.\nUses flat structure with format arrays instead of nested export configs.","properties":{"dir":{"default":"artifacts","description":"Output directory for all benchmark artifacts. Created if it doesn't exist.","format":"path","title":"Dir","type":"string"},"prefix":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Base filename override applied to ALL profile and server-metrics exports. With prefix='foo' every output becomes `foo.csv`, `foo.json`, `foo_timeslices.{csv,json}`, `foo.jsonl`, `foo_raw.jsonl`, `foo_outputs.json`, `foo_gpu_telemetry.jsonl`, `foo_server_metrics.{jsonl,json,csv,parquet}`. When unset (the default), historical per-file names are used: `profile_export_aiperf.csv/json`, `profile_export.jsonl`, `profile_export_raw.jsonl`, `outputs.json`, `gpu_telemetry_export.jsonl`, `server_metrics_export.{jsonl,json,csv,parquet}`. Known suffixes (`_raw.jsonl`, `_outputs.json`, `_timeslices.{csv,json}`, `_gpu_telemetry.jsonl`, `_server_metrics.{jsonl,json,csv,parquet}`, `.csv`/`.json`/`.jsonl`/`.parquet`) are stripped from the supplied value so `--profile-export-prefix foo_raw.jsonl` still yields a clean `foo` base.","title":"Prefix"},"summary":{"anyOf":[{"items":{"const":"json","type":"string"},"type":"array"},{"const":false,"type":"boolean"}],"description":"Summary export formats. Only 'json' is wired up to this field; the CSV summary is emitted regardless. Set to false to disable the summary JSON file only.","title":"Summary"},"records":{"anyOf":[{"items":{"const":"jsonl","type":"string"},"type":"array"},{"const":false,"type":"boolean"}],"description":"Per-request records export formats. Only 'jsonl' is wired up today. Set to false to disable the per-record JSONL file.","title":"Records"},"raw":{"default":false,"description":"Export raw request/response payloads as JSONL.","title":"Raw","type":"boolean"},"trace":{"default":false,"description":"Export HTTP trace data for debugging.","title":"Trace","type":"boolean"},"sliceDuration":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Time slice duration in seconds for trend analysis (must be > 0). Divides benchmark into windows for per-window statistics. Supports: 30, '30s', '5m', '2h'.","title":"Sliceduration","x-jinja2-supported":true},"showTraceTiming":{"default":false,"description":"Display HTTP trace timing metrics in console output. Shows detailed timing breakdown: blocked, DNS, connecting, sending, waiting (TTFB), receiving, and total duration.","title":"Showtracetiming","type":"boolean"},"userFiles":{"description":"User-defined templated files materialized into the run directory before the benchmark begins.","items":{"$ref":"#/$defs/UserFile"},"title":"Userfiles","type":"array"},"autoPlot":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Auto-invoke `aiperf plot` against the artifact directory after the benchmark completes. Unset (the default) means a `plot:` section implies True; set False to suppress plots even with one, True to force them. `AIPerfConfig` resolves this to a concrete bool, so any config that came through the loader carries a plain bool.","title":"Autoplot"},"plotRequired":{"default":false,"description":"Treat auto-plot failures as fatal: re-raise so `aiperf profile` exits non-zero. Only meaningful when auto_plot is True. Default False = warn and continue.","title":"Plotrequired","type":"boolean"},"exportOutputsJson":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Export generated response text after the run, to `outputs.json` or `_outputs.json` when `prefix` is set. Unset (the default) means the `raw` export level implies it; set False to opt out at raw, True to force it at any level. Always a concrete bool once validation has run.","title":"Exportoutputsjson"}},"title":"ArtifactsConfig","type":"object"},"AudioConfig":{"additionalProperties":false,"description":"Configuration for synthetic audio generation in multimodal datasets.\n\nControls the generation of synthetic audio for speech-to-text\nand audio-language model benchmarking. Generated audio is random\nnoise with specified sample rate, bit depth, and format.","properties":{"batchSize":{"default":0,"description":"Number of audio inputs to include in each multimodal request. Supported with chat endpoint type for multimodal models. Set to 0 to disable audio inputs.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of audio inputs to include in each multimodal request. Supported with chat endpoint type for multimodal models. Set to 0 to disable audio inputs.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"length":{"description":"Audio duration in seconds. Can be a fixed value or {mean, stddev} distribution. Used when batch_size > 0 for multimodal benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Length","x-jinja2-supported":true},"format":{"$ref":"#/$defs/AudioFormat","default":"wav","description":"File format for generated audio files. wav: uncompressed PCM (larger files). mp3: compressed (smaller files). Format affects file size in multimodal requests but not audio characteristics."},"sampleRates":{"description":"List of audio sample rates in kHz to randomly select from. Common values: 8.0 (telephony), 16.0 (speech), 44.1 (CD quality), 48.0 (professional). Specify multiple values for mixed-quality testing.","items":{"type":"number"},"minItems":1,"title":"Samplerates","type":"array"},"depths":{"description":"List of audio bit depths in bits to randomly select from. Each audio file is assigned a random depth from this list. Common values: 8 (low quality), 16 (CD quality), 24 (professional), 32 (high-end). Specify multiple values for mixed-quality testing.","items":{"type":"integer"},"minItems":1,"title":"Depths","type":"array"},"channels":{"default":1,"description":"Number of audio channels. 1 = mono (single channel), 2 = stereo (left/right channels). Stereo doubles file size. Most speech models use mono.","title":"Channels","oneOf":[{"default":1,"description":"Number of audio channels. 1 = mono (single channel), 2 = stereo (left/right channels). Stereo doubles file size. Most speech models use mono.","maximum":2,"minimum":1,"title":"Channels","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"AudioConfig","type":"object"},"AudioFormat":{"description":"Audio file formats for synthetic audio generation.","enum":["wav","mp3"],"title":"AudioFormat","type":"string"},"BenchmarkConfig":{"additionalProperties":false,"description":"Pure runtime configuration - what SystemController and services need.\n\nContains all fields required to execute a single benchmark run.\nDoes NOT include sweep, multi_run, variables, or random_seed settings\n(those live on AIPerfConfig as envelope-level cross-variation fields).\n\nRequired Sections:\n - models: Model(s) to benchmark\n - endpoint: Server connection settings\n - datasets: Single-element list of dataset configuration\n - phases: Ordered list of benchmark phases. A single-dict under\n ``phases:`` is rejected; use the envelope-level ``warmup:`` /\n ``profiling:`` shorthand or an explicit list.\n\nOptional Sections:\n - artifacts: Export and console settings\n - slos: SLO-based quality metrics (generic dict)\n - tokenizer: Token counting configuration\n - gpu_telemetry: GPU metrics from DCGM endpoints\n - server_metrics: Server metrics from Prometheus endpoints\n - runtime: Worker and communication settings\n - logging: Logging and debug settings\n - metrics: Metrics aggregation configuration\n - accuracy: Accuracy evaluation configuration\n\nValidation:\n The configuration is validated in several stages:\n 1. Individual field validation (types, ranges, formats)\n 2. Cross-field validation (mutual exclusivity, dependencies)\n\nEnvironment Variables:\n Values can reference environment variables using ${VAR} syntax.\n Optional defaults: ${VAR:default_value}\n\n Example:\n api_key: ${OPENAI_API_KEY}\n api_key: ${OPENAI_API_KEY:sk-default}","properties":{"models":{"description":"Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced. Accepts a model name string, list of model names, or full ModelsAdvanced config.","oneOf":[{"type":"string","description":"Single model name (normalized to ModelsAdvanced)."},{"type":"array","items":{"type":"string"},"minItems":1,"description":"List of model names (normalized to ModelsAdvanced)."},{"$ref":"#/$defs/ModelsAdvanced"}]},"endpoint":{"$ref":"#/$defs/EndpointConfig","description":"Endpoint configuration for connecting to inference server(s). Includes URLs, API type, authentication, timeout, and connection settings."},"datasets":{"description":"Dataset configuration as a single-element list. The list shape exists to share the schema between YAML and the AIPerfSweep CRD; the runtime currently loads exactly one dataset. Singular `dataset:` shorthand at the BenchmarkConfig top level is normalized to a one-entry list with name='default'.","items":{"discriminator":{"propertyName":"type","mapping":{"synthetic":"#/$defs/SyntheticDataset","file":"#/$defs/FileDataset","public":"#/$defs/PublicDataset"}},"anyOf":[{"$ref":"#/$defs/SyntheticDataset"},{"$ref":"#/$defs/FileDataset"},{"$ref":"#/$defs/PublicDataset"}]},"maxItems":1,"minItems":1,"title":"Datasets","type":"array"},"phases":{"description":"Ordered benchmark phases. Each entry must have a unique 'name' (e.g. 'warmup', 'profiling'). Order in the list IS the execution order; the first phase runs first. Single-config shorthand ({'type': 'concurrency', ...}) is normalized to a list of one. Top-level 'warmup:'/'profiling:' shorthand is normalized to a [warmup, profiling] list pre-validation. Shorthand: use a single phase config directly (with 'type' key) and it is normalized to a one-entry phases list.","oneOf":[{"description":"Ordered benchmark phases. Each entry must have a unique 'name' (e.g. 'warmup', 'profiling'). Order in the list IS the execution order; the first phase runs first. Single-config shorthand ({'type': 'concurrency', ...}) is normalized to a list of one. Top-level 'warmup:'/'profiling:' shorthand is normalized to a [warmup, profiling] list pre-validation.","items":{"discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}},"anyOf":[{"$ref":"#/$defs/ConcurrencyPhase"},{"$ref":"#/$defs/PoissonPhase"},{"$ref":"#/$defs/GammaPhase"},{"$ref":"#/$defs/ConstantPhase"},{"$ref":"#/$defs/UserCentricPhase"},{"$ref":"#/$defs/FixedSchedulePhase"}]},"minItems":1,"title":"Phases","type":"array"},{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Single phase config (normalized to a one-entry phases list).","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}}]},"artifacts":{"$ref":"#/$defs/ArtifactsConfig","description":"Artifacts configuration for benchmark output. Controls output directory and export formats."},"slos":{"anyOf":[{"additionalProperties":{"oneOf":[{"type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"type":"object"},{"type":"null"}],"default":null,"description":"SLO (Service Level Objectives) configuration as a generic dict. Maps metric names to threshold values. A request is counted as good only if it meets ALL specified thresholds.","title":"Slos","x-jinja2-supported":true},"tokenizer":{"anyOf":[{"$ref":"#/$defs/TokenizerConfig"},{"type":"null"}],"default":null,"description":"Tokenizer configuration for token counting. Used for ISL/OSL enforcement and accurate metrics. If not specified, uses the first model name."},"gpuTelemetry":{"$ref":"#/$defs/GpuTelemetryConfig","description":"GPU telemetry configuration for DCGM metrics collection. Collects GPU metrics (power, utilization, temperature) from DCGM endpoints. Enabled by default. Set enabled: false to disable."},"serverMetrics":{"$ref":"#/$defs/ServerMetricsConfig","description":"Server metrics configuration for Prometheus scraping. Collects operational metrics (queue depth, KV cache, batch sizes) from inference server Prometheus endpoints. Enabled by default. Set enabled: false to disable."},"networkLatency":{"$ref":"#/$defs/NetworkLatencyConfig","description":"Network latency calibration configuration. Probes the endpoint RTT during profiling and subtracts the mean from latency metrics, emitted as separate network_adjusted_* metrics. Disabled by default."},"otel":{"$ref":"#/$defs/OTelConfig"},"mlflow":{"$ref":"#/$defs/MLflowConfig"},"wandb":{"$ref":"#/$defs/WandbConfig"},"runtime":{"$ref":"#/$defs/RuntimeConfig","description":"Runtime configuration for worker processes and inter-process communication."},"logging":{"$ref":"#/$defs/LoggingConfig"},"metrics":{"$ref":"#/$defs/MetricsConfig","description":"Metrics aggregation configuration for benchmark summaries."},"accuracy":{"anyOf":[{"$ref":"#/$defs/AccuracyConfig"},{"type":"null"}],"default":null,"description":"Accuracy benchmarking configuration. When set, enables accuracy evaluation alongside performance profiling."},"scenario":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Lock all benchmark invariants for a named scenario (e.g. 'inferencex-agentx-mvp'). Plain data here; the lock is applied by the ScenarioResolver step in the pre-bootstrap resolver chain (auto-fills defaults, validates, raises ScenarioLockError on conflict). Distinct from the sweep ``scenarios`` strategy (SweepConfig), which expands hand-picked named runs -- this field is a single invariant LOCK, not a sweep.","title":"Scenario"},"unsafeOverride":{"default":false,"description":"Convert scenario lock errors to warnings; stamps submission_valid=false in the resolved scenario outcome. No-op without ``scenario``. Plain data; consumed by the ScenarioResolver.","title":"Unsafeoverride","type":"boolean"},"model":{"description":"Shorthand for 'models'. Accepts a single value which will be converted to a list/dict. Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced.","oneOf":[{"type":"string"},{"type":"array","items":{"type":"string"}},{"$ref":"#/$defs/ModelsAdvanced","description":"Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced.","x-kubernetes-preserve-unknown-fields":true}],"x-singular-alias-of":"models"},"dataset":{"anyOf":[{"additionalProperties":false,"description":"Synthetic dataset configuration.\n\nGenerates prompts programmatically based on token length\nspecifications. Ideal for controlled experiments.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"synthetic","description":"Dataset type discriminator. Must be 'synthetic'.","title":"Type","type":"string"},"entries":{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","minimum":1,"title":"Entries","type":"integer"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic dataset generation. When set, makes synthetic prompts, sampling, and other random operations reproducible across runs. Essential for A/B testing and debugging. Overrides global random_seed for this dataset.","title":"Randomseed"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptConfig"},{"type":"null"}],"default":null,"description":"Prompt/token length configuration specifying ISL, OSL, sequence distributions, and batch processing settings."},"prefixPrompts":{"anyOf":[{"$ref":"#/$defs/PrefixPromptConfig"},{"type":"null"}],"default":null,"description":"Shared prefix configuration for KV cache testing. Generates prefix prompts that are prepended to user prompts, simulating cached context scenarios."},"turns":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Number of request-response turns per conversation. Can be a fixed integer or {mean, stddev} distribution. Each turn consists of a user message and model response. Set to 1 for single-turn interactions. Multi-turn conversations enable testing of context retention and conversation history handling.","title":"Turns"},"turnDelay":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Delay in milliseconds between consecutive turns within a multi-turn conversation. Can be a fixed value or {mean, stddev} distribution. Simulates user think time between receiving a response and sending the next message. Only applies when turns > 1. Set to 0 for back-to-back turns.","title":"Turndelay"},"turnDelayRatio":{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","minimum":0.0,"title":"Turndelayratio","type":"number"},"images":{"anyOf":[{"$ref":"#/$defs/ImageConfig"},{"type":"null"}],"default":null,"description":"Synthetic image configuration for multimodal vision-language testing."},"audio":{"anyOf":[{"$ref":"#/$defs/AudioConfig"},{"type":"null"}],"default":null,"description":"Synthetic audio configuration for multimodal speech/audio testing."},"video":{"anyOf":[{"$ref":"#/$defs/VideoConfig"},{"type":"null"}],"default":null,"description":"Synthetic video configuration for multimodal video understanding testing."},"rankings":{"anyOf":[{"$ref":"#/$defs/RankingsConfig"},{"type":"null"}],"default":null,"description":"Rankings/reranking configuration for generating query-passage pairs. Only relevant for rankings endpoint types."},"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.isl; hoisted into prompts before validation.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.osl; hoisted into prompts before validation.","title":"Osl"}},"required":["type"],"title":"SyntheticDataset","type":"object"},{"additionalProperties":false,"description":"File-based dataset configuration.\n\nLoads prompts from a local file in various formats.\nSupports trace replay and custom sampling strategies.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"file","description":"Dataset type discriminator. Must be 'file'.","title":"Type","type":"string"},"path":{"anyOf":[{"format":"path","type":"string"},{"$ref":"#/$defs/RemoteRef"},{"type":"null"}],"default":null,"description":"Path to file or directory containing benchmark dataset, or a remote reference: `hf://[datasets/]/[@][/]`, `s3:///[?versionId=]`, or an `https://`/`http://` URL. Can be absolute or relative. Mutually exclusive with `records:`. Supported formats depend on the format field: JSONL for single_turn/multi_turn, JSONL (optionally gzipped) for tracelab, JSONL trace files for mooncake_trace/bailian_trace, Parquet for baseten_trace, directories for random_pool.","title":"Path"},"records":{"anyOf":[{"items":{"additionalProperties":true,"type":"object"},"type":"array"},{"additionalProperties":{"items":{"additionalProperties":true,"type":"object"},"type":"array"},"type":"object"},{"type":"null"}],"default":null,"description":"Inline benchmark records, embedded directly in the YAML config. Mutually exclusive with `path:`. The element schema is determined by `format:` (same shape as one line of the equivalent JSONL file). For `format: random_pool`, may be either a flat list (single pool) or a dict-of-lists (multi-pool, mirrors the directory-of-JSONLs file mode). All other formats require a flat list.","title":"Records"},"format":{"$ref":"#/$defs/DatasetFormat","default":"single_turn","description":"Dataset file format determining parsing logic and expected file structure. single_turn: JSONL with single prompt-response exchanges. multi_turn: JSONL with conversation history. tracelab: TraceLab agentic-coding corpus, JSONL or gzipped JSONL. mooncake_trace / bailian_trace / baseten_trace / burst_gpt_trace: timestamped trace files for replay. sagemaker_data_capture: JSONL captured by SageMaker DataCapture. random_pool: directory of reusable prompts."},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration. Allows scaling timestamps and token lengths before replay. Applies to trace formats such as mooncake_trace and baseten_trace, except speedup_ratio, which is rejected for baseten_trace (use replay_speedup / --replay-speedup to scale replay pacing there)."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Limit number of records to use from file. If not specified, uses all records in the file.","title":"Entries"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic sampling. When set, makes random/shuffle sampling reproducible across runs. Overrides global random_seed for this dataset.","title":"Randomseed"},"traceSessionSampleRatio":{"anyOf":[{"exclusiveMinimum":0.0,"maximum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Fraction of trace sessions to keep for replay, sampled whole-session to preserve multi-turn integrity; deterministic when ``random_seed`` is set. Only supported by the baseten_trace loader.","title":"Tracesessionsampleratio"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds; ``None`` disables the cap. Honored by the DAG JSONL loader and the baseten_trace loader's closed-loop think-times; the clamp count is reported at end of load.","title":"Interturndelaycapseconds"},"maxIdleGapCapSeconds":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Collapse idle gaps between consecutive requests (across all sessions) to at most this many seconds, so a sparse or session-sampled trace does not replay dead air; ``None`` disables the cap. The cap is in replay wall-clock seconds, applied after replay_speedup compression, so it bounds actual benchmark idle time regardless of speedup. Only supported by the baseten_trace loader.","title":"Maxidlegapcapseconds"},"replaySpeedup":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Trace replay wall-clock compression (10 = 10x faster than recorded): divides normalized timestamps and inter-turn delays; ``None`` = real time. Unlike synthesis speedup_ratio, hash_ids stay untouched (KV-cache fidelity). Only supported by the baseten_trace loader.","title":"Replayspeedup"},"openLoopReplay":{"default":true,"description":"Open-loop replay (the default): each session starts at its absolute, speedup-scaled recorded timestamp; continuation turns fire at max(recorded timestamp, prior-turn completion). Set ``False`` for closed-loop back-pressure: continuation turns fire a think-time (recorded start-to-start gap minus recorded e2e duration) after the prior turn completes, keeping sessions causally ordered when replayed service times differ from recorded (e.g. A/A comparisons). Only honored by the baseten_trace loader.","title":"Openloopreplay","type":"boolean"},"openLoopStrict":{"default":false,"description":"In open-loop replay, fire every trace row at its absolute recorded timestamp as an independent single-turn session, trading away multi-turn grouping and session metrics. Only honored by the baseten_trace loader.","title":"Openloopstrict","type":"boolean"},"omitKvHints":{"default":false,"description":"Drop recorded KV-cache hints (``hash_ids``, ``block_size``) from replayed request bodies, for strict frontends that reject unknown parameters. Only honored by the baseten_trace loader.","title":"Omitkvhints","type":"boolean"},"forceMinTokens":{"default":true,"description":"Pin ``min_tokens`` to the recorded output length so replayed generations match recorded lengths; disable to let EOS end generations naturally (some servers reject ``min_tokens``). Only honored by the baseten_trace loader.","title":"Forcemintokens","type":"boolean"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one. Can be a fixed integer or {mean, stddev} distribution. Per-line `output_length` values in the file always take precedence.","title":"Osl"},"promptBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of text items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable text inputs entirely (e.g. image/audio/video-only workloads).","title":"Promptbatchsize"},"imageBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of images per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable image inputs entirely.","title":"Imagebatchsize"},"audioBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of audio items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable audio inputs entirely.","title":"Audiobatchsize"},"videoBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of video items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable video inputs entirely.","title":"Videobatchsize"},"blockSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Hash-id block granularity for trace replay (--isl-block-size). hash-id trace loaders (mooncake_trace, bailian_trace, baseten_trace, tracelab) decode each hash_id into a cached block of this many tokens; total ISL = (num_hash_ids - 1) * block_size + final_block_size. When unset, the loader's plugin-metadata default applies (e.g. 512 for mooncake_trace, 16 for bailian_trace). Not used by weka, which carries its own inline per-block sizes.","title":"Blocksize"},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree. The idle clock covers initial future work and restarts when the final in-flight request across the root and all descendant streams completes. If no request from that tree reaches the wire before the cap, AIPerf uniformly advances only that tree's pending replay timers. Dataset timestamps are unchanged, and request order, spawn/join dependencies, and replay barriers remain authoritative. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds"},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from trace datasets at load time. With this flag, Turn.timestamp and Turn.delay are emitted as None so concurrency / request-rate timing modes dispatch turns back-to-back instead of reproducing the recorded user think-time gaps. No effect under fixed-schedule (timestamps drive that mode before they could be ignored -- combine with --no-fixed-schedule if you want both behaviors). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For weka_trace inputs, emit Turn.delay using only the recorded per-request `think_time` (client-side delay before each request) instead of the full `t_curr - t_prev` inter-request delta. Compresses replay wall time against zero-latency mocks because the recorded `api_time` portion of each gap is dropped. Mirrors kv-cache-tester's default `--timing-strategy think-only`. Falls back to the full delta for turns whose recorded `think_time` is null. Mutually exclusive with ignore_trace_delays. No effect on non-weka trace loaders.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per Weka root trace. Only honored by format weka_trace: the loader drops whole traces whose *recorded* peak exceeds this ceiling at load time (filter-then-cap before entries). Not a DatasetManager tokenize filter; rejected for non-Weka formats.","title":"Maxcontextlength"},"remoteFixedScheduleCheck":{"anyOf":[{"$ref":"#/$defs/RemoteFixedScheduleCheck"},{"type":"null"}],"default":null,"description":"Internal: set by the CLI for a remote `--input-file` and cleared by the dataset preflight once the first record decides fixed-schedule auto-promotion. Do not set in YAML."},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for trace replay (mirror of the synthetic dataset's prompts.cache_bust). The engine reads the active dataset's target via ``BenchmarkConfig.get_cache_bust_target()``; 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."}},"required":["type"],"title":"FileDataset","type":"object"},{"additionalProperties":false,"description":"Public dataset configuration.\n\nUses well-known public benchmarking datasets that are\nautomatically downloaded and processed by AIPerf.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"public","description":"Dataset type discriminator. Must be 'public'.","title":"Type","type":"string"},"dataset":{"$ref":"#/$defs/PublicDatasetType","description":"Pre-configured public dataset to download and use for benchmarking. Name of the HuggingFace public dataset enum (e.g. 'sharegpt', 'alpaca'). AIPerf automatically downloads and parses these datasets."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Limit number of records to use from the dataset. If not specified, uses all available records.","title":"Entries"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic sampling from the dataset. Overrides global random_seed for this dataset.","title":"Randomseed"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"hfSubset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset subset/config name override (e.g. 'sharegpt4o'). Only applies for HuggingFace-backed public dataset loaders. Takes priority over the subset defined in the plugin registry.","title":"Hfsubset"},"filters":{"additionalProperties":{"type":"string"},"description":"Dataset-specific filters forwarded to public dataset loaders. Supported keys and values depend on the selected dataset.","title":"Filters","type":"object"},"hfMooncakeDataset":{"anyOf":[{"minLength":1,"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace repository for mooncake_hf; required for this loader. The CLI flag --hf-mooncake-dataset auto-selects this loader.","title":"Hfmooncakedataset"},"hfMooncakeSplit":{"anyOf":[{"minLength":1,"pattern":"\\S","type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace split override for mooncake_hf (e.g. train or test). Unset uses the plugin default, test. Only supported by mooncake_hf.","title":"Hfmooncakesplit"},"mooncakeTraceToHfMapping":{"additionalProperties":{"type":"string"},"description":"Mooncake target field to dotted source column path for mooncake_hf. Supports messages, session_id, tools, output_length, extra, and extra.PARAMETER.","title":"Mooncaketracetohfmapping","type":"object"},"hfWekaDataset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset repo for the generic Weka loader (e.g. `semianalysisai/cc-traces-weka-061526`). Only applies with `dataset: weka_hf`; setting it with any other public dataset is an error. Passing `--hf-weka-dataset` on the CLI auto-selects `--public-dataset weka_hf`, so the repo flag works on its own. Pinned Weka public dataset aliases keep their registry-defined repo names.","title":"Hfwekadataset"},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds for HF-backed Weka trace replay (mirror of ``FileDataset.inter_turn_delay_cap_seconds``). ``None`` disables the cap.","title":"Interturndelaycapseconds"},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree for HF-backed Weka replay (mirror of ``FileDataset.trace_idle_gap_cap_seconds``). Dataset timestamps remain unchanged. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds"},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from HF-backed Weka trace replay at load time (mirror of ``FileDataset.ignore_trace_delays``). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For HF-backed Weka replay, emit Turn.delay using only the recorded per-request ``think_time`` instead of the full inter-request delta (mirror of ``FileDataset.use_think_time_only``). Mutually exclusive with ignore_trace_delays.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per conversation for HF-backed Weka replay; drops over-length traces at load using *recorded* lengths (mirror of ``FileDataset.max_context_length``). Rejected for non-Weka public datasets.","title":"Maxcontextlength"},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for HF-backed Weka trace replay (mirror of ``FileDataset.cache_bust``). 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration for HF-backed Weka trace replay (mirror of ``FileDataset.synthesis``). ``--max-isl``/``--max-osl`` cap the per-conversation input/output lengths of the replayed HF Weka traces; without this they applied only to file-based traces. ``None`` disables synthesis."},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one (mirror of ``FileDataset.osl``). Can be a fixed integer or {mean, stddev} distribution. Per-record output lengths always take precedence.","title":"Osl"}},"required":["type","dataset"],"title":"PublicDataset","type":"object"}],"description":"Single dataset config (normalized to a one-entry datasets list).","x-singular-alias-of":"datasets"},"warmup":{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Warmup phase shorthand. Requires profiling when used.","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}},"profiling":{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Profiling phase shorthand. Normalized to a phases entry named 'profiling'.","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}}},"required":["endpoint"],"title":"BenchmarkConfig","type":"object","allOf":[{"anyOf":[{"required":["model"]},{"required":["models"]}]},{"anyOf":[{"required":["dataset"]},{"required":["datasets"]}]},{"anyOf":[{"required":["phases"]},{"required":["profiling"]}]}]},"CacheBustConfig":{"additionalProperties":false,"description":"Per-run cache-bust marker configuration.\n\nTwo families of targets:\n\nRID targets (``system_prefix``, ``system_suffix``, ``first_turn_prefix``,\n``first_turn_suffix``): inject a per-trajectory unique marker derived from a\nSHA-256 digest of (benchmark_id, recycle_pass, trajectory_index, trace_id).\nThe digest is phase-agnostic — warmup and profiling turns for the same\ntrajectory share the same marker — so warmup KV-cache work transfers to\nprofiling. Different trajectories (different lanes or recycle passes) receive\ndistinct markers. Prefix variants diverge at token 0; suffix variants append\nafter existing content.\n\nWarmup-isolation targets (``warmup_isolation_system``,\n``warmup_isolation_first_turn``): inject a constant ``[warmup]`` marker only\nduring the WARMUP phase; during PROFILING the marker is ``None`` so the\nprofiling request arrives clean. These targets are phase-aware and do NOT use\na per-trajectory digest. Incompatible with ``agentic_replay`` timing mode.\n\n``media_uuids`` extends a RID target to media: the ``uuid`` of every media\ncontent part gets the trajectory's digest as a suffix, so a replayed\ntrajectory cannot reuse the server's multimodal cache entries for the same\nmedia either.","properties":{"target":{"$ref":"#/$defs/CacheBustTarget","default":"none","description":"Where (and how) to inject a per-conversation cache-bust marker. Prefix variants prepend at token 0 (most aggressive); suffix variants append after existing content. 'none' disables the feature (default)."},"mediaUuids":{"default":false,"description":"Append the trajectory's RID digest to every media `uuid`, so a replay cannot hit vLLM's multimodal cache entries from an earlier play. Requires a RID target and the chat endpoint.","title":"Mediauuids","type":"boolean"}},"title":"CacheBustConfig","type":"object"},"CacheBustTarget":{"description":"Where (and how) to inject a per-conversation cache-bust marker.\n\nPrefix variants diverge at token 0 of the prompt (most aggressive -- defeats\nKV-cache prefix matching for the entire prompt). Suffix variants append\nafter existing content (preserves leading-prefix caching).","enum":["none","system_prefix","system_suffix","first_turn_prefix","first_turn_suffix","warmup_isolation_system","warmup_isolation_first_turn"],"title":"CacheBustTarget","type":"string"},"CancellationConfig":{"additionalProperties":false,"description":"Configuration for request cancellation testing.\n\nEnables testing server behavior when clients cancel requests mid-flight.","properties":{"rate":{"description":"Percentage of requests to cancel (0-100). 0.5 means half a percent; 10.0 means 10%% of requests will be cancelled.","title":"Rate","oneOf":[{"description":"Percentage of requests to cancel (0-100). 0.5 means half a percent; 10.0 means 10%% of requests will be cancelled.","maximum":100.0,"minimum":0.0,"title":"Rate","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"delay":{"default":0.0,"description":"Seconds to wait after sending before cancelling. 0.0 means cancel immediately after send.","title":"Delay","oneOf":[{"default":0.0,"description":"Seconds to wait after sending before cancelling. 0.0 means cancel immediately after send.","minimum":0.0,"title":"Delay","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["rate"],"title":"CancellationConfig","type":"object"},"ConcurrencyPhase":{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","title":"Concurrency","oneOf":[{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"ConcurrencyPhase","type":"object"},"ConnectionReuseStrategy":{"description":"Transport connection reuse strategy. Controls how and when connections are reused across requests.","enum":["pooled","never","sticky-user-sessions"],"title":"ConnectionReuseStrategy","type":"string"},"ConstantPhase":{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"ConvergenceConfig":{"additionalProperties":false,"description":"Adaptive trial-stopping criterion.\n\nPresence of this object on `MultiRunConfig.convergence` enables\nadaptive stopping: trials run until the criterion fires (or\n`MultiRunConfig.num_runs` is reached, whichever comes first).","properties":{"metric":{"description":"Metric tag to evaluate convergence on (e.g., 'ttft').","title":"Metric","type":"string"},"stat":{"$ref":"#/$defs/ConvergenceStat","default":"avg","description":"Statistic on the metric to inspect (avg/p50/p90/p95/p99)."},"mode":{"$ref":"#/$defs/ConvergenceCriterionType","default":"ci_width","description":"Convergence criterion plugin name (ci_width|cv|distribution)."},"threshold":{"anyOf":[{"exclusiveMaximum":1,"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Criterion threshold (interpretation depends on `mode`). When None (default), each criterion uses its own algorithm-specific default: ci_width=0.10, cv=0.05, distribution=0.05 (KS p-value).","title":"Threshold","x-jinja2-supported":true},"minRuns":{"default":2,"description":"Minimum trials before the criterion is checked.","title":"Minruns","oneOf":[{"default":2,"description":"Minimum trials before the criterion is checked.","minimum":2,"title":"Minruns","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metric"],"title":"ConvergenceConfig","type":"object"},"ConvergenceCriterionType":{"enum":["ci_width","cv","distribution"],"title":"ConvergenceCriterionType","type":"string"},"ConvergenceStat":{"description":"Statistic to evaluate for convergence when using ci_width or cv mode.","enum":["avg","p50","p90","p95","p99","min","max"],"title":"ConvergenceStat","type":"string"},"DatasetFormat":{"enum":["burst_gpt_trace","bailian_trace","baseten_trace","mooncake_trace","h_cua_perf","raw_payload","inputs_json","dag_jsonl","sagemaker_data_capture","multi_turn","random_pool","single_turn","speed_bench_qualitative","speed_bench_throughput_1k","speed_bench_throughput_2k","speed_bench_throughput_8k","speed_bench_throughput_16k","speed_bench_throughput_32k","speed_bench_throughput_100k","tracelab","weka_trace"],"title":"DatasetFormat","type":"string"},"DatasetSamplingStrategy":{"enum":["random","sequential","shuffle"],"title":"DatasetSamplingStrategy","type":"string"},"DualBindCommunicationConfig":{"additionalProperties":false,"description":"Dual-bind (IPC + TCP) communication configuration.\n\nFor Kubernetes deployments where controller services connect via IPC\n(co-located in same pod) and worker pods connect via TCP.\n\nWhen controller_host is None, services use IPC (local mode).\nWhen controller_host is set, services use TCP to that host (remote mode).","properties":{"type":{"const":"dual","description":"Communication type. Must be 'dual'.","title":"Type","type":"string"},"ipcPath":{"default":"/tmp/aiperf","description":"Directory for IPC socket files (local services).","title":"Ipcpath","type":"string"},"tcpHost":{"default":"0.0.0.0","description":"TCP bind host for proxies. Use 0.0.0.0 to listen on all interfaces.","title":"Tcphost","type":"string"},"controllerHost":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Controller host for remote TCP connections. When set, services connect via TCP to this host instead of IPC. In Kubernetes, set via JobSet DNS (e.g., controller.namespace.svc).","title":"Controllerhost"},"recordsPort":{"default":5557,"description":"Port for records push/pull communication.","title":"Recordsport","oneOf":[{"default":5557,"description":"Port for records push/pull communication.","maximum":65535,"minimum":1,"title":"Recordsport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditRouterPort":{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","title":"Creditrouterport","oneOf":[{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Creditrouterport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"controlPort":{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","title":"Controlport","oneOf":[{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Controlport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditReturnPushPullPort":{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","title":"Creditreturnpushpullport","oneOf":[{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","maximum":65535,"minimum":1,"title":"Creditreturnpushpullport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"eventBusProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Event bus proxy ports (XPUB/XSUB)."},"datasetManagerProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Dataset manager proxy ports (DEALER/ROUTER)."},"rawInferenceProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Raw inference proxy ports (PUSH/PULL)."}},"required":["type"],"title":"DualBindCommunicationConfig","type":"object"},"EmpiricalDistribution":{"additionalProperties":false,"description":"Discrete distribution sampled from weighted values.\n\nYAML:\n isl:\n points:\n - {value: 128, weight: 40}\n - {value: 512, weight: 35}\n - {value: 2048, weight: 20}\n - {value: 8192, weight: 5}","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"points":{"description":"Weighted discrete values. Weights are relative.","items":{"$ref":"#/$defs/EmpiricalPoint"},"title":"Points","type":"array"},"type":{"const":"empirical","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["points"],"title":"EmpiricalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"EmpiricalPoint":{"additionalProperties":false,"description":"A weighted value in an empirical distribution.","properties":{"value":{"description":"The discrete value.","title":"Value","oneOf":[{"description":"The discrete value.","title":"Value","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"weight":{"default":1.0,"description":"Relative weight (normalized internally).","title":"Weight","oneOf":[{"default":1.0,"description":"Relative weight (normalized internally).","exclusiveMinimum":0.0,"title":"Weight","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["value"],"title":"EmpiricalPoint","type":"object"},"EndpointConfig":{"additionalProperties":false,"description":"Endpoint configuration for connecting to inference servers.\n\nThis section configures how AIPerf connects to and communicates\nwith the target inference server(s). It supports multiple URLs\nfor load-balanced deployments and various API types.","properties":{"urls":{"description":"List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","items":{"type":"string"},"title":"Urls","type":"array","x-kubernetes-preserve-unknown-fields":true},"urlStrategy":{"$ref":"#/$defs/URLSelectionStrategy","default":"round_robin","description":"Strategy for distributing requests across multiple URLs. round_robin cycles through URLs in order."},"type":{"$ref":"#/$defs/EndpointType","default":"chat","description":"API endpoint type determining request/response format. chat: OpenAI chat completions, completions: OpenAI completions, embeddings: vector embeddings, rankings: reranking, template: custom format, and others — see `aiperf plugins` for the full list."},"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Override default endpoint path. Use for servers with non-standard API paths. Example: '/custom/v2/generate'","title":"Path"},"apiKey":{"anyOf":[{"type":"string"},{"type":"null"}],"description":"API authentication key. Supports environment variable substitution: ${OPENAI_API_KEY}. Can also use ${VAR:default} syntax for defaults.","title":"Apikey"},"timeout":{"default":21600,"description":"Request timeout in seconds (0 = no timeout). Requests exceeding this duration are marked as failed. Should exceed expected max response time.","title":"Timeout","oneOf":[{"default":21600,"description":"Request timeout in seconds (0 = no timeout). Requests exceeding this duration are marked as failed. Should exceed expected max response time.","minimum":0.0,"title":"Timeout","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"streaming":{"default":false,"description":"Enable streaming (Server-Sent Events) responses. Required for accurate TTFT (time to first token) measurement. Server must support streaming for this to work.","title":"Streaming","type":"boolean"},"allowEmptyContent":{"default":false,"description":"Retain explicitly empty content or reasoning strings as timed responses for streaming OpenAI Chat Completions. This changes the shared response timeline without adding output tokens.","title":"Allowemptycontent","type":"boolean"},"transport":{"anyOf":[{"$ref":"#/$defs/TransportType"},{"type":"null"}],"default":null,"description":"Transport plugin name. 'http' (aiohttp-based HTTP/1.1) and 'sagemaker' (AWS SageMaker Runtime, normally derived from sagemaker.endpoint_name) are shipped. Auto-detected from the URL when unset, which always selects 'http'; explicit setting overrides auto-detection."},"awsRegion":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"AWS region for the request. Required when auth_type='sigv4'.","title":"Awsregion"},"awsProfile":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Named AWS credentials profile. Unset uses botocore's default credential chain.","title":"Awsprofile"},"authType":{"anyOf":[{"$ref":"#/$defs/RequestSignerType"},{"type":"null"}],"default":null,"description":"Request signing method for authentication. When set, the selected request_signer plugin signs every HTTP request sent by the HTTP transport. Replaces Bearer token auth (api_key is ignored when auth_type is set)."},"awsService":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SigV4 signing name -- the credential scope the signature is bound to (e.g. 'execute-api', 'sagemaker', 'bedrock'). This is not always the API id: the 'sagemaker-runtime' API signs as 'sagemaker' and 'bedrock-runtime' as 'bedrock'. Required when auth_type='sigv4', unless the selected transport declares which AWS API it speaks, in which case the scope is resolved from botocore's service model.","title":"Awsservice"},"sagemaker":{"$ref":"#/$defs/SageMakerConfig","description":"SageMaker Runtime routing options. Setting sagemaker.endpoint_name selects the SageMaker transport."},"connectionReuse":{"$ref":"#/$defs/ConnectionReuseStrategy","default":"pooled","description":"HTTP connection management strategy. pooled: shared connection pool (fastest), never: new connection per request (includes TCP overhead), sticky-user-sessions: dedicated connection per user session."},"useLegacyMaxTokens":{"default":false,"description":"Use 'max_tokens' field instead of 'max_completion_tokens'. Enable for compatibility with older OpenAI API versions.","title":"Uselegacymaxtokens","type":"boolean"},"useServerTokenCount":{"default":false,"description":"Use server-reported token counts from response usage field. When true, trusts usage.prompt_tokens and usage.completion_tokens. When false, counts tokens locally using configured tokenizer.","title":"Useservertokencount","type":"boolean"},"perChunkUsage":{"default":false,"description":"Request per-chunk token usage on streaming responses by setting stream_options.continuous_usage_stats. When true, the server reports cumulative usage on every chunk (not just the final one), which lets inter-token latency subtract the first content chunk's real token count instead of assuming one token. Requires a server that supports continuous_usage_stats (e.g. vLLM, TRT-LLM); strict OpenAI rejects it. Requires use_server_token_count so OSL and the first-chunk count share the server source.","title":"Perchunkusage","type":"boolean"},"template":{"anyOf":[{"$ref":"#/$defs/TemplateConfig"},{"type":"null"}],"default":null,"description":"Custom template configuration for template endpoint type. Only used when type='template'. Defines request body format and response parsing."},"headers":{"additionalProperties":{"type":"string"},"description":"Custom HTTP headers to include in all requests. Useful for authentication, tracing, or routing. Values support environment variable substitution.","title":"Headers","type":"object"},"extra":{"additionalProperties":true,"description":"Additional fields to include in request body. Merged into every request. Common fields: temperature, top_p, top_k, stop.","title":"Extra","type":"object"},"downloadVideoContent":{"default":false,"description":"For video generation endpoints, download the video content after generation completes. Adds a content download step to the async polling flow.","title":"Downloadvideocontent","type":"boolean"},"requestContentType":{"anyOf":[{"$ref":"#/$defs/RequestContentType"},{"type":"null"}],"default":null,"description":"Content type for request body serialization. Default is 'application/json'. Set to 'multipart/form-data' for servers that require form-encoded requests (e.g. vLLM video generation)."},"sessionHeader":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HTTP header name used to carry the per-session affinity identifier. When set, replaces the default `X-Correlation-ID` header. Useful when the inference server expects a custom session-affinity header (e.g. `--session-header X-Session-ID`).","title":"Sessionheader"},"uuidAndStrip":{"default":false,"description":"Enable AIPerf-managed image stripping for vLLM's multimodal processor cache. Dataset-authored image UUIDs, including UUID-only cache references, always pass through on the chat endpoint; this flag only strips repeated content after AIPerf observes it in the same session. Automatic stripping supports only single_turn datasets with session_id-grouped rows; multi_turn is rejected. The server cache must cover the working set, and requests in a session must reach a replica that retains earlier UUIDs.","title":"Uuidandstrip","type":"boolean"},"waitForModelTimeout":{"default":0.0,"description":"Enable a pre-flight readiness probe by setting this to a positive value (seconds). aiperf applies this timeout to each URL/model probe before starting the benchmark, aborting with a non-zero exit if any probe times out. For multiple URLs or models, worst-case wall-clock time can be roughly this timeout multiplied by the number of URL/model probes. The probe strategy is controlled by `--wait-for-model-mode`, which defaults to sending a 1-token inference request. 0 (default) disables the probe. Eliminates the need for external shell-based readiness loops in containers and Kubernetes recipes.","title":"Waitformodeltimeout","oneOf":[{"default":0.0,"description":"Enable a pre-flight readiness probe by setting this to a positive value (seconds). aiperf applies this timeout to each URL/model probe before starting the benchmark, aborting with a non-zero exit if any probe times out. For multiple URLs or models, worst-case wall-clock time can be roughly this timeout multiplied by the number of URL/model probes. The probe strategy is controlled by `--wait-for-model-mode`, which defaults to sending a 1-token inference request. 0 (default) disables the probe. Eliminates the need for external shell-based readiness loops in containers and Kubernetes recipes.","minimum":0.0,"title":"Waitformodeltimeout","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"waitForModelInterval":{"default":5.0,"description":"Seconds between readiness probe attempts. Only consulted when `--wait-for-model-timeout` is positive.","title":"Waitformodelinterval","oneOf":[{"default":5.0,"description":"Seconds between readiness probe attempts. Only consulted when `--wait-for-model-timeout` is positive.","exclusiveMinimum":0.0,"title":"Waitformodelinterval","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"waitForModelMode":{"default":"inference","description":"Strategy for the readiness probe. 'inference' (default): POST a 1-token inference request to the configured endpoint; this is the strongest signal — it proves the full stack (frontend, scheduler, worker, forward pass) is live. Any HTTP status < 500 counts as ready. 'models': GET `/v1/models` and verify the model id appears in `data[]` (cheaper, no tokens consumed; falls back to a plain GET on the base URL on 404). 'both': run 'models' first, then 'inference'. Only consulted when `--wait-for-model-timeout` is positive.","enum":["models","inference","both"],"title":"Waitformodelmode","type":"string"},"resetKvCache":{"anyOf":[{"type":"boolean"},{"$ref":"#/$defs/ResetKvCacheConfig"},{"type":"null"}],"default":null,"description":"When enabled, POST a KV-cache reset once per logical benchmark cell before warmup/profiling. Accepts false | true | {timeout_seconds?, path?}."},"serverProfiler":{"anyOf":[{"type":"boolean"},{"$ref":"#/$defs/ServerProfilerConfig"},{"type":"null"}],"default":null,"description":"When enabled, start/stop the server profiler around each profiling phase. Accepts false | true | {timeout_seconds?, start_path?, stop_path?}."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","oneOf":[{"type":"string"},{"description":"List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","items":{"type":"string"},"title":"Urls","type":"array","x-kubernetes-preserve-unknown-fields":true}],"x-singular-alias-of":"urls"}},"required":[],"title":"EndpointConfig","type":"object","allOf":[{"anyOf":[{"required":["url"]},{"required":["urls"]}]}]},"EndpointType":{"enum":["audio_transcription","chat","cohere_rankings","completions","responses","messages","chat_embeddings","embeddings","hf_tei_rankings","huggingface_generate","image_generation","image_edit","video_generation","image_retrieval","nim_embeddings","nim_rankings","solido_rag","raw","template"],"title":"EndpointType","type":"string"},"FileDataset":{"additionalProperties":false,"description":"File-based dataset configuration.\n\nLoads prompts from a local file in various formats.\nSupports trace replay and custom sampling strategies.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"file","description":"Dataset type discriminator. Must be 'file'.","title":"Type","type":"string"},"path":{"anyOf":[{"format":"path","type":"string"},{"$ref":"#/$defs/RemoteRef"},{"type":"null"}],"default":null,"description":"Path to file or directory containing benchmark dataset, or a remote reference: `hf://[datasets/]/[@][/]`, `s3:///[?versionId=]`, or an `https://`/`http://` URL. Can be absolute or relative. Mutually exclusive with `records:`. Supported formats depend on the format field: JSONL for single_turn/multi_turn, JSONL (optionally gzipped) for tracelab, JSONL trace files for mooncake_trace/bailian_trace, Parquet for baseten_trace, directories for random_pool.","title":"Path"},"records":{"anyOf":[{"items":{"additionalProperties":true,"type":"object"},"type":"array"},{"additionalProperties":{"items":{"additionalProperties":true,"type":"object"},"type":"array"},"type":"object"},{"type":"null"}],"default":null,"description":"Inline benchmark records, embedded directly in the YAML config. Mutually exclusive with `path:`. The element schema is determined by `format:` (same shape as one line of the equivalent JSONL file). For `format: random_pool`, may be either a flat list (single pool) or a dict-of-lists (multi-pool, mirrors the directory-of-JSONLs file mode). All other formats require a flat list.","title":"Records"},"format":{"$ref":"#/$defs/DatasetFormat","default":"single_turn","description":"Dataset file format determining parsing logic and expected file structure. single_turn: JSONL with single prompt-response exchanges. multi_turn: JSONL with conversation history. tracelab: TraceLab agentic-coding corpus, JSONL or gzipped JSONL. mooncake_trace / bailian_trace / baseten_trace / burst_gpt_trace: timestamped trace files for replay. sagemaker_data_capture: JSONL captured by SageMaker DataCapture. random_pool: directory of reusable prompts."},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration. Allows scaling timestamps and token lengths before replay. Applies to trace formats such as mooncake_trace and baseten_trace, except speedup_ratio, which is rejected for baseten_trace (use replay_speedup / --replay-speedup to scale replay pacing there)."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Limit number of records to use from file. If not specified, uses all records in the file.","title":"Entries","x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic sampling. When set, makes random/shuffle sampling reproducible across runs. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"traceSessionSampleRatio":{"anyOf":[{"exclusiveMinimum":0.0,"maximum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Fraction of trace sessions to keep for replay, sampled whole-session to preserve multi-turn integrity; deterministic when ``random_seed`` is set. Only supported by the baseten_trace loader.","title":"Tracesessionsampleratio","x-jinja2-supported":true},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds; ``None`` disables the cap. Honored by the DAG JSONL loader and the baseten_trace loader's closed-loop think-times; the clamp count is reported at end of load.","title":"Interturndelaycapseconds","x-jinja2-supported":true},"maxIdleGapCapSeconds":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Collapse idle gaps between consecutive requests (across all sessions) to at most this many seconds, so a sparse or session-sampled trace does not replay dead air; ``None`` disables the cap. The cap is in replay wall-clock seconds, applied after replay_speedup compression, so it bounds actual benchmark idle time regardless of speedup. Only supported by the baseten_trace loader.","title":"Maxidlegapcapseconds","x-jinja2-supported":true},"replaySpeedup":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Trace replay wall-clock compression (10 = 10x faster than recorded): divides normalized timestamps and inter-turn delays; ``None`` = real time. Unlike synthesis speedup_ratio, hash_ids stay untouched (KV-cache fidelity). Only supported by the baseten_trace loader.","title":"Replayspeedup","x-jinja2-supported":true},"openLoopReplay":{"default":true,"description":"Open-loop replay (the default): each session starts at its absolute, speedup-scaled recorded timestamp; continuation turns fire at max(recorded timestamp, prior-turn completion). Set ``False`` for closed-loop back-pressure: continuation turns fire a think-time (recorded start-to-start gap minus recorded e2e duration) after the prior turn completes, keeping sessions causally ordered when replayed service times differ from recorded (e.g. A/A comparisons). Only honored by the baseten_trace loader.","title":"Openloopreplay","type":"boolean"},"openLoopStrict":{"default":false,"description":"In open-loop replay, fire every trace row at its absolute recorded timestamp as an independent single-turn session, trading away multi-turn grouping and session metrics. Only honored by the baseten_trace loader.","title":"Openloopstrict","type":"boolean"},"omitKvHints":{"default":false,"description":"Drop recorded KV-cache hints (``hash_ids``, ``block_size``) from replayed request bodies, for strict frontends that reject unknown parameters. Only honored by the baseten_trace loader.","title":"Omitkvhints","type":"boolean"},"forceMinTokens":{"default":true,"description":"Pin ``min_tokens`` to the recorded output length so replayed generations match recorded lengths; disable to let EOS end generations naturally (some servers reject ``min_tokens``). Only honored by the baseten_trace loader.","title":"Forcemintokens","type":"boolean"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one. Can be a fixed integer or {mean, stddev} distribution. Per-line `output_length` values in the file always take precedence.","title":"Osl"},"promptBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of text items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable text inputs entirely (e.g. image/audio/video-only workloads).","title":"Promptbatchsize","x-jinja2-supported":true},"imageBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of images per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable image inputs entirely.","title":"Imagebatchsize","x-jinja2-supported":true},"audioBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of audio items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable audio inputs entirely.","title":"Audiobatchsize","x-jinja2-supported":true},"videoBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of video items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable video inputs entirely.","title":"Videobatchsize","x-jinja2-supported":true},"blockSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hash-id block granularity for trace replay (--isl-block-size). hash-id trace loaders (mooncake_trace, bailian_trace, baseten_trace, tracelab) decode each hash_id into a cached block of this many tokens; total ISL = (num_hash_ids - 1) * block_size + final_block_size. When unset, the loader's plugin-metadata default applies (e.g. 512 for mooncake_trace, 16 for bailian_trace). Not used by weka, which carries its own inline per-block sizes.","title":"Blocksize","x-jinja2-supported":true},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree. The idle clock covers initial future work and restarts when the final in-flight request across the root and all descendant streams completes. If no request from that tree reaches the wire before the cap, AIPerf uniformly advances only that tree's pending replay timers. Dataset timestamps are unchanged, and request order, spawn/join dependencies, and replay barriers remain authoritative. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds","x-jinja2-supported":true},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from trace datasets at load time. With this flag, Turn.timestamp and Turn.delay are emitted as None so concurrency / request-rate timing modes dispatch turns back-to-back instead of reproducing the recorded user think-time gaps. No effect under fixed-schedule (timestamps drive that mode before they could be ignored -- combine with --no-fixed-schedule if you want both behaviors). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For weka_trace inputs, emit Turn.delay using only the recorded per-request `think_time` (client-side delay before each request) instead of the full `t_curr - t_prev` inter-request delta. Compresses replay wall time against zero-latency mocks because the recorded `api_time` portion of each gap is dropped. Mirrors kv-cache-tester's default `--timing-strategy think-only`. Falls back to the full delta for turns whose recorded `think_time` is null. Mutually exclusive with ignore_trace_delays. No effect on non-weka trace loaders.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per Weka root trace. Only honored by format weka_trace: the loader drops whole traces whose *recorded* peak exceeds this ceiling at load time (filter-then-cap before entries). Not a DatasetManager tokenize filter; rejected for non-Weka formats.","title":"Maxcontextlength","x-jinja2-supported":true},"remoteFixedScheduleCheck":{"anyOf":[{"$ref":"#/$defs/RemoteFixedScheduleCheck"},{"type":"null"}],"default":null,"description":"Internal: set by the CLI for a remote `--input-file` and cleared by the dataset preflight once the first record decides fixed-schedule auto-promotion. Do not set in YAML."},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for trace replay (mirror of the synthetic dataset's prompts.cache_bust). The engine reads the active dataset's target via ``BenchmarkConfig.get_cache_bust_target()``; 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."}},"required":["name","type"],"title":"FileDataset","type":"object"},"FixedDistribution":{"title":"FixedDistribution","description":"Returns a constant value on every sample. Scalars coerce to this. Can be a number (interpreted as a fixed value) or {value} object.","anyOf":[{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable resolving to a fixed distribution value."},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template resolving to a fixed distribution value."},{"additionalProperties":false,"description":"Returns a constant value on every sample. Scalars coerce to this.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"value":{"description":"The constant value returned on every sample.","title":"Value","type":"number"},"type":{"const":"fixed","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["value"],"title":"FixedDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"integer"},{"type":"number"}]},"FixedSchedulePhase":{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset","x-jinja2-supported":true},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"FixedSchedulePhase","type":"object"},"GPUTelemetryCollectorType":{"enum":["dcgm","pynvml","amdsmi"],"title":"GPUTelemetryCollectorType","type":"string"},"GPUTelemetryMode":{"description":"GPU telemetry display mode.","enum":["summary","realtime_dashboard"],"title":"GPUTelemetryMode","type":"string"},"GammaPhase":{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"GpuTelemetryConfig":{"title":"GpuTelemetryConfig","description":"GPU telemetry configuration for live or replayed GPU metrics collection.\n\nCollects GPU metrics through DCGM exporter endpoints by default. The\n``collector`` field can switch to a local backend (``pynvml`` for NVIDIA,\n``amdsmi`` for AMD ROCm); ``mode`` controls summary vs. realtime dashboard\ndisplay.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9400/metrics\"\n → GpuTelemetryConfig(enabled=True, urls=[\"http://localhost:9400/metrics\"])\n - Singular url field: {url: \"...\"}\n → GpuTelemetryConfig(urls=[\"...\"]) Can be a URL string or full config object.","anyOf":[{"additionalProperties":false,"description":"GPU telemetry configuration for live or replayed GPU metrics collection.\n\nCollects GPU metrics through DCGM exporter endpoints by default. The\n``collector`` field can switch to a local backend (``pynvml`` for NVIDIA,\n``amdsmi`` for AMD ROCm); ``mode`` controls summary vs. realtime dashboard\ndisplay.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9400/metrics\"\n → GpuTelemetryConfig(enabled=True, urls=[\"http://localhost:9400/metrics\"])\n - Singular url field: {url: \"...\"}\n → GpuTelemetryConfig(urls=[\"...\"])","properties":{"enabled":{"default":true,"description":"Enable GPU telemetry collection. Set to false to disable.","title":"Enabled","type":"boolean"},"urls":{"description":"DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","items":{"type":"string"},"title":"Urls","type":"array"},"metricsFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to CSV file with pre-recorded GPU metrics. Alternative to live DCGM collection.","title":"Metricsfile"},"collector":{"$ref":"#/$defs/GPUTelemetryCollectorType","default":"dcgm","description":"GPU telemetry collector backend. Use 'dcgm' for DCGM exporter endpoints or a local collector (e.g. 'pynvml' for NVIDIA, 'amdsmi' for AMD ROCm) for on-host metrics collection."},"mode":{"$ref":"#/$defs/GPUTelemetryMode","default":"summary","description":"GPU telemetry display mode. Summary emits aggregate console output; realtime_dashboard enables live dashboard updates."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","oneOf":[{"type":"string"},{"description":"DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","items":{"type":"string"},"title":"Urls","type":"array"}],"x-singular-alias-of":"urls"}},"title":"GpuTelemetryConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"string"}]},"GridSweep":{"additionalProperties":false,"description":"Grid sweep - all combinations of parameters (Cartesian product).","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"grid","default":"grid","description":"Sweep type discriminator.","title":"Type","type":"string"},"parameters":{"additionalProperties":{"items":{},"type":"array"},"description":"Parameters to sweep: dot-notation path -> list of values.","minProperties":1,"title":"Parameters","type":"object"}},"required":["parameters"],"title":"GridSweep","type":"object"},"ImageConfig":{"additionalProperties":false,"allOf":[{"if":{"properties":{"source":{"const":"noise"}}},"then":{"properties":{"sourceSampling":{"const":"random-with-replacement"}}}}],"description":"Configuration for synthetic image generation in multimodal datasets.\n\nControls the generation of synthetic images for vision-language\nmodel benchmarking. Images are generated by randomly sampling and\nresizing source images to specified dimensions.","properties":{"batchSize":{"default":0,"description":"Number of images to include in each multimodal request. Supported with chat endpoint type for vision-language models. Set to 0 to disable image inputs. Higher batch sizes test multi-image understanding and increase request payload size.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of images to include in each multimodal request. Supported with chat endpoint type for vision-language models. Set to 0 to disable image inputs. Higher batch sizes test multi-image understanding and increase request payload size.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"width":{"description":"Image width in pixels. Can be a fixed integer or {mean, stddev} distribution. Combined with height to determine image dimensions and file sizes for multimodal benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Width","x-jinja2-supported":true},"height":{"description":"Image height in pixels. Can be a fixed integer or {mean, stddev} distribution. Used when batch_size > 0 for multimodal vision benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Height","x-jinja2-supported":true},"format":{"$ref":"#/$defs/ImageFormat","default":"jpeg","description":"Image file format for generated images. png: lossless compression (larger files, best quality). jpeg: lossy compression (smaller files, good quality). random: randomly select between PNG and JPEG per image. Format affects file size in multimodal requests and encoding overhead."},"source":{"anyOf":[{"$ref":"#/$defs/ImageSource"},{"format":"path","type":"string"}],"default":"noise","description":"Source image generation mode (default: noise). noise: generate random noise images on the fly — no files on disk, effectively unbounded variety so servers cannot dedupe identical inputs. assets: index images from the bundled assets/source_images directory and lazily load them at the requested dimensions. A path string indexes images from the given directory (e.g. ./source_images). Random-noise images are roughly incompressible, so payload bytes are larger than equivalent natural images.","title":"Source"},"sourceSampling":{"$ref":"#/$defs/ImageSourceSamplingStrategy","default":"random-with-replacement","description":"How source images are selected from finite image sources selected by source='assets' or a directory path. random-with-replacement: draw each source image independently; repeats may occur immediately. shuffle-cycle: draw every source image once per shuffled cycle, reshuffling after exhaustion. sequential-cycle: walk source images in sorted load order and wrap after exhaustion. For noise mode, only random-with-replacement is valid because there is no finite source pool."}},"title":"ImageConfig","type":"object"},"ImageFormat":{"description":"Image file formats for synthetic image generation.","enum":["png","jpeg","random"],"title":"ImageFormat","type":"string"},"ImageSource":{"description":"Source image generation mode for multimodal benchmarking.","enum":["assets","noise"],"title":"ImageSource","type":"string"},"ImageSourceSamplingStrategy":{"description":"How source images are selected from a finite source-image pool.","enum":["random-with-replacement","shuffle-cycle","sequential-cycle"],"title":"ImageSourceSamplingStrategy","type":"string"},"IpcCommunicationConfig":{"additionalProperties":false,"description":"IPC (Unix socket) communication configuration.\n\nFor single-machine deployments with lowest latency.\nUses Unix domain sockets for all inter-service communication.","properties":{"type":{"const":"ipc","description":"Communication type. Must be 'ipc'.","title":"Type","type":"string"},"path":{"default":"/tmp/aiperf","description":"Directory for IPC socket files. AIPerf creates multiple socket files in this directory.","title":"Path","type":"string"}},"required":["type"],"title":"IpcCommunicationConfig","type":"object"},"JsonValue":{},"LatinHypercubeSweep":{"additionalProperties":false,"description":"Latin hypercube sampling — perfect marginal stratification per axis.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"samples":{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","title":"Samples","oneOf":[{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","maximum":1048576,"minimum":2,"title":"Samples","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"seed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"RNG seed for reproducibility. Must be non-negative; scipy's QMC engines reject negative seeds with an opaque low-level error.","title":"Seed","x-jinja2-supported":true},"dimensions":{"description":"Dimensions to sample over.","items":{"$ref":"#/$defs/SamplingDimension"},"minItems":1,"title":"Dimensions","type":"array"},"labelFormat":{"default":"index","description":"Variation-label format. 'index' => '_NNNN' (short, sortable). 'kv' => 'k1=v1, k2=v2' (grid-style).","enum":["index","kv"],"title":"Labelformat","type":"string"},"type":{"const":"latin_hypercube","default":"latin_hypercube","description":"Sweep type discriminator.","title":"Type","type":"string"},"optimization":{"default":"random-cd","description":"LHS post-processing to reduce centered discrepancy. None = vanilla LHS.","enum":[null,"random-cd","lloyd"],"title":"Optimization"}},"required":["samples","dimensions"],"title":"LatinHypercubeSweep","type":"object"},"LogNormalDistribution":{"additionalProperties":false,"description":"Log-normal parameterized by mean and median (right-skewed, always positive).\n\nSkew is controlled by the mean/median ratio: larger ratio = heavier right tail.\nWhen mean == median the distribution is deterministic.\n\nInternally: sigma = sqrt(2 * log(mean / median)), mu = log(median).","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"mean":{"description":"Desired mean of the output distribution.","title":"Mean","oneOf":[{"description":"Desired mean of the output distribution.","exclusiveMinimum":0.0,"title":"Mean","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"median":{"description":"Desired median. Must be <= mean. Lower median = more right skew.","title":"Median","oneOf":[{"description":"Desired median. Must be <= mean. Lower median = more right skew.","exclusiveMinimum":0.0,"title":"Median","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"type":{"const":"lognormal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["mean","median"],"title":"LogNormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"LoggingConfig":{"additionalProperties":false,"description":"Logging configuration for verbosity and debug settings.","properties":{"level":{"$ref":"#/$defs/AIPerfLogLevel","default":"INFO","description":"Global logging verbosity level (loguru-style severity ladder; TRACE most verbose, CRITICAL least). NOTICE and SUCCESS are loguru-specific intermediate levels (between INFO and WARNING) and have no equivalent in the Python stdlib `logging` module — pick INFO/WARNING/ERROR if you want stdlib-compatible severities."}},"title":"LoggingConfig","type":"object"},"MLflowConfig":{"additionalProperties":false,"description":"MLflow tracking and artifact-upload configuration.","properties":{"trackingUri":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"MLflow tracking URI.","title":"Trackinguri"},"experiment":{"default":"aiperf","description":"MLflow experiment name.","title":"Experiment","type":"string"},"runName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"MLflow run name.","title":"Runname"},"tags":{"anyOf":[{"type":"array","items":{"oneOf":[{"type":"string"},{"type":"array","minItems":2,"maxItems":2}]}},{"type":"string"},{"additionalProperties":{"type":"string"},"type":"object"},{"type":"null"}],"default":null,"description":"Additional MLflow run tags to attach on upload.","title":"Tags"},"parentRunId":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Optional MLflow parent run ID.","title":"Parentrunid"},"artifactGlobs":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Artifact glob overrides for MLflow upload.","title":"Artifactglobs"}},"title":"MLflowConfig","type":"object"},"MetricsConfig":{"additionalProperties":false,"description":"Configuration for benchmark metric aggregation behavior.\n\nCurrently empty — list-valued record metrics aggregate via t-digest at\nfirst-touch (see :mod:`aiperf.metrics.list_metric_aggregation`); the\nsketch compression is tunable via ``AIPERF_METRICS_TDIGEST_COMPRESSION``.\nReserved as the home for any future user-facing metric-aggregation knobs.","properties":{},"title":"MetricsConfig","type":"object"},"ModelItem":{"additionalProperties":false,"description":"Configuration for a single model in advanced models configuration.\n\nUsed when the models section uses the advanced format with\nexplicit items, weights, and per-model settings.","properties":{"name":{"description":"Model name or identifier as known to the inference server.","minLength":1,"title":"Name","type":"string"},"weight":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Selection weight for weighted strategy (0.0-1.0). Weights must sum to 1.0 (+/-0.01) across all models; they are validated, not auto-normalized. Example: weight=0.7 means ~70%% of requests to this model.","title":"Weight","x-jinja2-supported":true},"lora":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"LoRA adapter name to load with this model. Server must support dynamic LoRA adapter loading.","title":"Lora"},"modalities":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"List of input modalities this model supports. [Currently a no-op: no selection strategy consumes this field. Accepted for forward-compatibility / declarative documentation.]","title":"Modalities"},"tokenizer":{"anyOf":[{"$ref":"#/$defs/TokenizerOverride"},{"type":"null"}],"default":null,"description":"Per-model tokenizer override. Use when this model requires a different tokenizer than global config."}},"required":["name"],"title":"ModelItem","type":"object"},"ModelSelectionStrategy":{"description":"Strategy for selecting the model to use for the request.","enum":["round_robin","random","weighted"],"title":"ModelSelectionStrategy","type":"string"},"ModelsAdvanced":{"additionalProperties":false,"description":"Advanced models configuration with selection strategy and item details.\n\nUse this format when you need weighted routing, LoRA adapters,\nor per-model tokenizer overrides.","properties":{"strategy":{"$ref":"#/$defs/ModelSelectionStrategy","default":"round_robin","description":"Strategy for selecting models when multiple are configured. round_robin cycles through models, random selects randomly, weighted uses configured weights."},"items":{"description":"List of model configurations. At least one model required.","items":{"$ref":"#/$defs/ModelItem"},"minItems":1,"title":"Items","type":"array"}},"required":["items"],"title":"ModelsAdvanced","type":"object"},"MultiRunConfig":{"additionalProperties":false,"description":"Trial mechanics within a single sweep variation.\n\nWhen `num_runs > 1` (or `convergence` is set), AIPerf runs multiple\ntrials per variation and aggregates statistics. When `num_runs == 1`\nand `convergence is None`, exactly one trial runs per variation.","properties":{"numRuns":{"default":1,"description":"Upper bound on trials per variation. When `convergence` is set, trials may stop earlier; this is the hard ceiling. Cap matches `BenchmarkPlan.trials` so programmatic plan assembly cannot exceed what user-facing config validation allows.","title":"Numruns","oneOf":[{"default":1,"description":"Upper bound on trials per variation. When `convergence` is set, trials may stop earlier; this is the hard ceiling. Cap matches `BenchmarkPlan.trials` so programmatic plan assembly cannot exceed what user-facing config validation allows.","maximum":10,"minimum":1,"title":"Numruns","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"cooldownSeconds":{"default":0.0,"description":"Cooldown between trials within a single variation. Capped at 24h to surface typos like `1e18` at config-load time rather than blocking the orchestrator inside ``asyncio.sleep``.","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between trials within a single variation. Capped at 24h to surface typos like `1e18` at config-load time rather than blocking the orchestrator inside ``asyncio.sleep``.","maximum":86400,"minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"confidenceLevel":{"default":0.95,"description":"Confidence level for aggregate statistics (0.90, 0.95, 0.99).","title":"Confidencelevel","oneOf":[{"default":0.95,"description":"Confidence level for aggregate statistics (0.90, 0.95, 0.99).","exclusiveMaximum":1,"exclusiveMinimum":0,"title":"Confidencelevel","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"setConsistentSeed":{"default":true,"description":"Auto-set a random seed when none is given for cross-trial consistency.","title":"Setconsistentseed","type":"boolean"},"varySeedPerTrial":{"default":false,"description":"When True, derive a distinct seed for each trial of a variation via SHA-256 over `(envelope_seed, variation.label, trial)`. When False (default), all trials of a variation share the variation's seed. Off gives pure-runtime variance for confidence statistics; on captures end-to-end variance including per-trial workload sampling.","title":"Varyseedpertrial","type":"boolean"},"disableWarmupAfterFirst":{"default":true,"description":"Skip warmup on trials 2..N of a variation (steady-state measurement).","title":"Disablewarmupafterfirst","type":"boolean"},"convergence":{"anyOf":[{"$ref":"#/$defs/ConvergenceConfig"},{"type":"null"}],"default":null,"description":"Optional adaptive early-stop. Absent = fixed-N (run `num_runs` trials)."}},"title":"MultiRunConfig","type":"object"},"MultimodalDistribution":{"additionalProperties":false,"description":"Weighted mixture of N peaks (N >= 2).\n\nYAML:\n isl:\n peaks:\n - {mean: 128, stddev: 20, weight: 60}\n - {mean: 2048, median: 1800, weight: 40}\n # Equal split — omit weight:\n isl:\n peaks:\n - {mean: 128, stddev: 20}\n - {mean: 2048, median: 1800}\n - {mean: 8192, median: 4096}","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"peaks":{"description":"Two or more weighted sub-distributions.","items":{"$ref":"#/$defs/PeakEntry"},"minItems":2,"title":"Peaks","type":"array"},"type":{"const":"multimodal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["peaks"],"title":"MultimodalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"NetworkLatencyConfig":{"additionalProperties":false,"description":"Network latency calibration configuration.\n\nWhen ``enabled``, AIPerf opens a fresh TCP connection to the endpoint\nhost:port on an interval during profiling and records the handshake RTT.\nThe mean RTT is subtracted from the request-start-anchored latency metrics\n(request_latency, time_to_first_token, time_to_first_output_token) to produce\nnon-destructive ``network_adjusted_*``\nvariants. ``mean_ms`` supplies a fixed mean RTT and skips active probing.","properties":{"enabled":{"default":false,"description":"Enable network latency calibration. When true, AIPerf measures the endpoint RTT (or uses mean_ms) and subtracts it from latency metrics, emitted as separate network_adjusted_* metrics.","title":"Enabled","type":"boolean"},"pingInterval":{"description":"Seconds between TCP-handshake RTT probes during profiling. Ignored when mean_ms is set.","title":"Pinginterval","oneOf":[{"description":"Seconds between TCP-handshake RTT probes during profiling. Ignored when mean_ms is set.","exclusiveMinimum":0.0,"title":"Pinginterval","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"meanMs":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Fixed mean network RTT in milliseconds to subtract, bypassing active probing. When set, no probes are sent and this value is used directly.","title":"Meanms","x-jinja2-supported":true}},"title":"NetworkLatencyConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},"NormalDistribution":{"additionalProperties":false,"description":"Gaussian (truncated at 0) parameterized by mean and stddev.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"mean":{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","title":"Mean","oneOf":[{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","minimum":0.0,"title":"Mean","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"stddev":{"default":0.0,"description":"Standard deviation. 0 = deterministic.","title":"Stddev","oneOf":[{"default":0.0,"description":"Standard deviation. 0 = deterministic.","minimum":0.0,"title":"Stddev","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"type":{"const":"normal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["mean"],"title":"NormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"OTelConfig":{"additionalProperties":false,"description":"OpenTelemetry metrics streaming configuration.","properties":{"metricsUrl":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"OTLP/HTTP metrics endpoint URL.","title":"Metricsurl"},"streamMetricsEnabled":{"default":true,"description":"Stream metric records to OTel.","title":"Streammetricsenabled","type":"boolean"},"streamTimingEnabled":{"default":true,"description":"Stream timing records to OTel.","title":"Streamtimingenabled","type":"boolean"},"customResourceAttributes":{"additionalProperties":{"type":"string"},"description":"Custom OTel resource attributes.","title":"Customresourceattributes","type":"object"},"genAiProvider":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"GenAI semantic convention provider override.","title":"Genaiprovider"}},"title":"OTelConfig","type":"object"},"Objective":{"additionalProperties":false,"description":"Single optimization objective.\n\nMultiple ``Objective`` entries on an ``AdaptiveSearchSweep`` enable\nPareto Bayesian optimization (e.g. qlognehvi / qnehvi / qehvi). A\nlength-1 list is single-objective BO.","properties":{"metric":{"description":"Metric tag to optimize (e.g. 'output_token_throughput').","title":"Metric","type":"string"},"stat":{"default":"avg","description":"Statistic on the metric.","enum":["avg","p50","p90","p95","p99"],"title":"Stat","type":"string"},"direction":{"$ref":"#/$defs/OptimizationDirection","description":"MAXIMIZE or MINIMIZE."},"threshold":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Pareto reference point for hypervolume computation. Trials worse than this on this objective do not contribute to hypervolume. When None, auto-derived from the worst observed value among Sobol initial points. Ignored for single-objective runs.","title":"Threshold","x-jinja2-supported":true}},"required":["metric","direction"],"title":"Objective","type":"object"},"OptimizationDirection":{"description":"Direction of optimization for a metric.\n\nMAXIMIZE: Higher values are preferred (e.g. throughput, goodput).\nMINIMIZE: Lower values are preferred (e.g. latency, TTFT, p99).","enum":["maximize","minimize"],"title":"OptimizationDirection","type":"string"},"OutcomeConstraint":{"additionalProperties":false,"description":"Feasibility gate on a metric the optimizer is *not* optimizing.\n\nDistinct from ``Objective.threshold`` (Pareto reference point) and from\n``SLAFilter`` (post-hoc filter on benchmark eligibility). An outcome\nconstraint masks BoTorch acquisition: candidates predicted to violate\nit are downweighted.","properties":{"metric":{"description":"Metric tag to constrain.","title":"Metric","type":"string"},"op":{"description":"Comparison operator.","enum":["<=",">=","=="],"title":"Op","type":"string"},"bound":{"description":"Threshold value.","title":"Bound","oneOf":[{"description":"Threshold value.","title":"Bound","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metric","op","bound"],"title":"OutcomeConstraint","type":"object"},"PeakEntry":{"title":"PeakEntry","description":"A weighted component in a multimodal distribution.\n\nThe weight and distribution fields are written inline in YAML:\n {mean: 128, stddev: 20, weight: 60}\n\nThe ``weight`` key is extracted before the remaining fields are parsed\nas a SamplingDistribution. Defaults to 1.0 (equal split when omitted).","anyOf":[{"additionalProperties":false,"description":"A weighted component in a multimodal distribution.\n\nThe weight and distribution fields are written inline in YAML:\n {mean: 128, stddev: 20, weight: 60}\n\nThe ``weight`` key is extracted before the remaining fields are parsed\nas a SamplingDistribution. Defaults to 1.0 (equal split when omitted).","properties":{"distribution":{"description":"The sub-distribution for this peak.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}],"title":"Distribution"},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["distribution"],"title":"PeakEntry","type":"object"},{"additionalProperties":false,"description":"Inline FixedDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"value":{"description":"The constant value returned on every sample.","title":"Value","type":"number"},"type":{"const":"fixed","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["value"],"title":"FixedDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline NormalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"mean":{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","minimum":0.0,"title":"Mean","type":"number"},"stddev":{"default":0.0,"description":"Standard deviation. 0 = deterministic.","minimum":0.0,"title":"Stddev","type":"number"},"type":{"const":"normal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["mean"],"title":"NormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline LogNormalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"mean":{"description":"Desired mean of the output distribution.","exclusiveMinimum":0.0,"title":"Mean","type":"number"},"median":{"description":"Desired median. Must be <= mean. Lower median = more right skew.","exclusiveMinimum":0.0,"title":"Median","type":"number"},"type":{"const":"lognormal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["mean","median"],"title":"LogNormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline MultimodalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"peaks":{"description":"Two or more weighted sub-distributions.","items":{"$ref":"#/$defs/PeakEntry"},"minItems":2,"title":"Peaks","type":"array"},"type":{"const":"multimodal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["peaks"],"title":"MultimodalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline EmpiricalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"points":{"description":"Weighted discrete values. Weights are relative.","items":{"$ref":"#/$defs/EmpiricalPoint"},"title":"Points","type":"array"},"type":{"const":"empirical","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["points"],"title":"EmpiricalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true}]},"PoissonPhase":{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"PostProcessSpec":{"additionalProperties":false,"description":"Post-aggregation hook spec emitted by a search recipe.\n\nResolved through the ``search_recipe_post_process`` plugin category and\ninvoked by ``aggregate_sweep_and_export`` after per-variation aggregation.\n\nExample: ``PostProcessSpec(handler=\"ttft_sla_curve\", params={\"sla_ms\": 200},\noutput_filename=\"ttft_sla_curve.json\")``.","properties":{"handler":{"description":"Registered post-process handler name. Looked up in the ``search_recipe_post_process`` plugin category.","title":"Handler","type":"string"},"params":{"additionalProperties":true,"description":"Free-form parameters forwarded to the handler. Shape is handler-defined; Any is intentional because handlers ingest dynamic recipe-derived inputs.","title":"Params","type":"object"},"output_filename":{"description":"Filename (relative to ``sweep_aggregate/`` under the artifact dir) to write the handler's output to. Must end in ``.json`` and contain no path separators -- the hook joins it with ``sweep_aggregate/`` and refuses to escape that directory. NUL bytes and dot-only stems are also rejected.","title":"Output Filename","type":"string"}},"required":["handler","output_filename"],"title":"PostProcessSpec","type":"object"},"PrefixPromptConfig":{"additionalProperties":false,"description":"Configuration for prefix prompts (KV cache testing).\n\nPrefix prompts allow testing KV cache efficiency by generating\nrequests that share common prefixes. This simulates scenarios\nlike system prompts or shared context that can be cached.\n\nNote: pool_size/length are mutually exclusive with shared_system_length\nand user_context_length.","properties":{"poolSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of distinct prefix prompts to generate for KV cache testing. Each prefix is prepended to user prompts, simulating cached context scenarios. Prefixes are randomly selected from pool per request. Mutually exclusive with shared_system_length/user_context_length.","title":"Poolsize","x-jinja2-supported":true},"length":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Token length for each prefix prompt in the pool. Only used when pool_size is set. The prefix is concatenated to the body at the token level, so the final prompt contains exactly prefix length + input length tokens. Mutually exclusive with shared_system_length/user_context_length.","title":"Length","x-jinja2-supported":true},"sharedSystemLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Length of shared system prompt in tokens. This prompt is identical across all sessions and appears as a system message. First part of a two-part prefix structure with high cache hit rate expected. Mutually exclusive with pool_size/length.","title":"Sharedsystemlength","x-jinja2-supported":true},"userContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Length of per-session user context prompt in tokens. Each dataset entry gets a unique user context prompt. Second part of two-part prefix structure with lower cache hit rate expected. Mutually exclusive with pool_size/length.","title":"Usercontextlength","x-jinja2-supported":true}},"title":"PrefixPromptConfig","type":"object"},"PromptConfig":{"additionalProperties":false,"description":"Configuration for prompt/token specifications in synthetic datasets.\n\nThis is the core configuration for controlling input sequence length (ISL)\nand output sequence length (OSL) in synthetic data generation.","properties":{"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Input sequence length in tokens. Can be a fixed integer (e.g., 512) or distribution {mean: 512, stddev: 50}. AIPerf generates prompts with lengths following a normal distribution around the mean (±stddev). Ignored when sequence_distribution is specified.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length (max tokens to request via max_completion_tokens). Can be a fixed integer or distribution {mean, stddev}. Controls response length for synthetic datasets. When not set, the model determines output length. Ignored when sequence_distribution is specified.","title":"Osl"},"blockSize":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Token block size for hash-based prompt caching in mooncake_trace datasets. When hash_ids are provided in trace entries, prompts are divided into blocks of this size. Each hash_id maps to a cached block, enabling simulation of KV-cache sharing patterns from production workloads. Total prompt length = (num_hash_ids - 1) * block_size + final_block_size.","title":"Blocksize","x-jinja2-supported":true},"batchSize":{"default":1,"description":"Number of text inputs to include in each request for batch processing endpoints. Supported by embeddings and rankings endpoint types where models can process multiple inputs simultaneously. Set to 1 for single-input requests. Not applicable to chat or completions endpoints.","title":"Batchsize","oneOf":[{"default":1,"description":"Number of text inputs to include in each request for batch processing endpoints. Supported by embeddings and rankings endpoint types where models can process multiple inputs simultaneously. Set to 1 for single-input requests. Not applicable to chat or completions endpoints.","minimum":1,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker configuration. Injects a deterministic marker into prompts to defeat KV-cache prefix matching. Disabled by default (target='none')."},"corpus":{"anyOf":[{"$ref":"#/$defs/PromptCorpus"},{"type":"null"}],"default":null,"description":"Source corpus for synthetic prompt text generation. 'sonnet' uses Shakespeare sonnets. 'coding' uses realistic coding content (code, bash output, JSON, error tracebacks, git diffs). When unset, the active dataset loader's default applies (most loaders default to 'sonnet'; agentic-coding loaders such as weka_trace default to 'coding')."},"sequenceDistribution":{"anyOf":[{"items":{"$ref":"#/$defs/SequenceDistributionEntry"},"type":"array"},{"type":"null"}],"default":null,"description":"Distribution of (ISL, OSL) pairs with probabilities for mixed workload simulation. Each entry specifies {isl, osl, probability}. Probabilities are percentages (0-100) and must sum to 100. When specified, requests are sampled from this distribution instead of using isl/osl fields.","title":"Sequencedistribution"},"randomRangeRatio":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Sample ISL and OSL uniformly from a ratio-defined integer window around the configured means. Accepts a single float (applied to both ISL and OSL) or a JSON object {\"input\": 0.3, \"output\": 0.5} for independent values. Mutually exclusive with sequence_distribution. When a tokenizer is configured, the ISL mean is automatically reduced by tokenizer.num_special_tokens_to_add(pair=False) so --isl represents total server-side input tokens.","title":"Randomrangeratio"},"randomCorpusStyle":{"$ref":"#/$defs/RandomCorpusStyle","default":"vllm","description":"Sampling formula for random_range_ratio. vllm (default): symmetric window [floor(mean*(1-r)), ceil(mean*(1+r))], ratio in [0, 1). sglang: lower-bounded window [max(1, int(mean*r)), mean], ratio in [0, 1]."}},"title":"PromptConfig","type":"object"},"PromptCorpus":{"description":"Corpus used for synthetic prompt text generation.","enum":["sonnet","coding","random"],"title":"PromptCorpus","type":"string"},"PromptSelectionConfig":{"additionalProperties":false,"description":"Slim prompts block for file/public datasets (corpus selection only).","properties":{"corpus":{"anyOf":[{"$ref":"#/$defs/PromptCorpus"},{"type":"null"}],"default":null,"description":"Source corpus for synthesized prompt text. 'sonnet' uses Shakespeare sonnets. 'coding' uses realistic coding content. When unset, the active dataset loader's default applies. Honored only where content is synthesized (synthetic + hash/trace replay); verbatim loaders ignore it."}},"title":"PromptSelectionConfig","type":"object"},"PublicDataset":{"additionalProperties":false,"description":"Public dataset configuration.\n\nUses well-known public benchmarking datasets that are\nautomatically downloaded and processed by AIPerf.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"public","description":"Dataset type discriminator. Must be 'public'.","title":"Type","type":"string"},"dataset":{"$ref":"#/$defs/PublicDatasetType","description":"Pre-configured public dataset to download and use for benchmarking. Name of the HuggingFace public dataset enum (e.g. 'sharegpt', 'alpaca'). AIPerf automatically downloads and parses these datasets."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Limit number of records to use from the dataset. If not specified, uses all available records.","title":"Entries","x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic sampling from the dataset. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"hfSubset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset subset/config name override (e.g. 'sharegpt4o'). Only applies for HuggingFace-backed public dataset loaders. Takes priority over the subset defined in the plugin registry.","title":"Hfsubset"},"filters":{"additionalProperties":{"type":"string"},"description":"Dataset-specific filters forwarded to public dataset loaders. Supported keys and values depend on the selected dataset.","title":"Filters","type":"object"},"hfMooncakeDataset":{"anyOf":[{"minLength":1,"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace repository for mooncake_hf; required for this loader. The CLI flag --hf-mooncake-dataset auto-selects this loader.","title":"Hfmooncakedataset"},"hfMooncakeSplit":{"anyOf":[{"minLength":1,"pattern":"\\S","type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace split override for mooncake_hf (e.g. train or test). Unset uses the plugin default, test. Only supported by mooncake_hf.","title":"Hfmooncakesplit"},"mooncakeTraceToHfMapping":{"additionalProperties":{"type":"string"},"description":"Mooncake target field to dotted source column path for mooncake_hf. Supports messages, session_id, tools, output_length, extra, and extra.PARAMETER.","title":"Mooncaketracetohfmapping","type":"object"},"hfWekaDataset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset repo for the generic Weka loader (e.g. `semianalysisai/cc-traces-weka-061526`). Only applies with `dataset: weka_hf`; setting it with any other public dataset is an error. Passing `--hf-weka-dataset` on the CLI auto-selects `--public-dataset weka_hf`, so the repo flag works on its own. Pinned Weka public dataset aliases keep their registry-defined repo names.","title":"Hfwekadataset"},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds for HF-backed Weka trace replay (mirror of ``FileDataset.inter_turn_delay_cap_seconds``). ``None`` disables the cap.","title":"Interturndelaycapseconds","x-jinja2-supported":true},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree for HF-backed Weka replay (mirror of ``FileDataset.trace_idle_gap_cap_seconds``). Dataset timestamps remain unchanged. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds","x-jinja2-supported":true},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from HF-backed Weka trace replay at load time (mirror of ``FileDataset.ignore_trace_delays``). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For HF-backed Weka replay, emit Turn.delay using only the recorded per-request ``think_time`` instead of the full inter-request delta (mirror of ``FileDataset.use_think_time_only``). Mutually exclusive with ignore_trace_delays.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per conversation for HF-backed Weka replay; drops over-length traces at load using *recorded* lengths (mirror of ``FileDataset.max_context_length``). Rejected for non-Weka public datasets.","title":"Maxcontextlength","x-jinja2-supported":true},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for HF-backed Weka trace replay (mirror of ``FileDataset.cache_bust``). 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration for HF-backed Weka trace replay (mirror of ``FileDataset.synthesis``). ``--max-isl``/``--max-osl`` cap the per-conversation input/output lengths of the replayed HF Weka traces; without this they applied only to file-based traces. ``None`` disables synthesis."},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one (mirror of ``FileDataset.osl``). Can be a fixed integer or {mean, stddev} distribution. Per-record output lengths always take precedence.","title":"Osl"}},"required":["name","type","dataset"],"title":"PublicDataset","type":"object"},"PublicDatasetType":{"enum":["exgentic_v2","exgentic","sharegpt","aimo","mmstar","mmvu","vision_arena","llava_onevision","aimo_aime","aimo_numina_cot","aimo_numina_1_5","spec_bench","spec_al_gsm8k","spec_al_math500","spec_al_humaneval","spec_al_mbpp","spec_al_mtbench","instruct_coder","blazedit_5k","blazedit_10k","librispeech","voxpopuli","gigaspeech","ami","spgispeech","semianalysis_cc_traces_weka","semianalysis_cc_traces_weka_no_subagents","semianalysis_cc_traces_weka_with_subagents","semianalysis_cc_traces_weka_with_subagents_256k","semianalysis_cc_traces_weka_with_subagents_060226","semianalysis_cc_traces_weka_with_subagents_060226_256k","semianalysis_cc_traces_weka_with_subagents_060526","semianalysis_cc_traces_weka_with_subagents_060526_256k","semianalysis_cc_traces_weka_with_subagents_060826","semianalysis_cc_traces_weka_with_subagents_060826_256k","semianalysis_cc_traces_weka_061326","semianalysis_cc_traces_weka_061326_256k","semianalysis_cc_traces_weka_061526","semianalysis_cc_traces_weka_061526_256k","semianalysis_cc_traces_weka_062126","semianalysis_cc_traces_weka_062126_256k","weka_hf","mooncake_hf","speed_bench_qualitative","speed_bench_coding","speed_bench_humanities","speed_bench_math","speed_bench_multilingual","speed_bench_qa","speed_bench_rag","speed_bench_reasoning","speed_bench_roleplay","speed_bench_stem","speed_bench_summarization","speed_bench_writing","speed_bench_throughput_1k","speed_bench_throughput_1k_low_entropy","speed_bench_throughput_1k_mixed","speed_bench_throughput_1k_high_entropy","speed_bench_throughput_2k","speed_bench_throughput_2k_low_entropy","speed_bench_throughput_2k_mixed","speed_bench_throughput_2k_high_entropy","speed_bench_throughput_8k","speed_bench_throughput_8k_low_entropy","speed_bench_throughput_8k_mixed","speed_bench_throughput_8k_high_entropy","speed_bench_throughput_16k","speed_bench_throughput_16k_low_entropy","speed_bench_throughput_16k_mixed","speed_bench_throughput_16k_high_entropy","speed_bench_throughput_32k","speed_bench_throughput_32k_low_entropy","speed_bench_throughput_32k_mixed","speed_bench_throughput_32k_high_entropy","speed_bench_throughput_100k","speed_bench_throughput_100k_low_entropy","speed_bench_throughput_100k_mixed","speed_bench_throughput_100k_high_entropy","h_cua_perf"],"title":"PublicDatasetType","type":"string"},"RampConfig":{"title":"RampConfig","description":"Configuration for gradual value ramping.\n\nControls how a value (concurrency, rate, etc.) transitions from\nstart to target over time. Can be a duration (number in seconds or string like '30s', '5m', '2h') or {duration, strategy} object.","anyOf":[{"additionalProperties":false,"description":"Configuration for gradual value ramping.\n\nControls how a value (concurrency, rate, etc.) transitions from\nstart to target over time.","properties":{"duration":{"description":"Seconds to ramp from start to target value.","exclusiveMinimum":0.0,"title":"Duration","type":"number"},"strategy":{"$ref":"#/$defs/RampType","default":"linear","description":"Ramp curve shape: linear (constant rate), exponential (slow start, fast finish), poisson (stochastic with guaranteed completion)."}},"required":["duration"],"title":"RampConfig","type":"object"},{"type":"integer"},{"type":"number"},{"type":"string"}]},"RampType":{"enum":["exponential","linear","poisson"],"title":"RampType","type":"string"},"RandomCorpusStyle":{"description":"Benchmark style for RANDOM corpus generation.\n\nControls the full set of behaviors that vary between tools. Each style\nis a bundle of decisions that together reproduce the statistical\ndistribution of a specific benchmarking tool:\n\n.. list-table::\n :header-rows: 1\n :widths: 30 35 35\n\n * - Concern\n - VLLM\n - SGLANG\n * - Token pool\n - ``valid_token_ids`` (special tokens excluded via ``all_special_ids``)\n - ``all_token_ids`` (full ``range(vocab_size)``, no exclusion)\n * - BOS / special-token adjustment\n - ``max(0, mean - num_special_tokens)``, applied to the mean before\n the window bounds are computed\n - ``max(1, drawn - num_special_tokens)``, applied per-request after\n sampling so the window keeps its raw-mean shape\n * - ISL/OSL range formula\n - Symmetric: ``[floor(mean*(1-r)), ceil(mean*(1+r))]``\n - Lower-bounded: ``[max(1, int(mean*r)), mean]``\n * - RNG algorithm (preseed)\n - PCG64 via ``numpy.random.default_rng(seed)``\n - MT19937 via a private ``numpy.random.RandomState(seed)``\n * - RNG draw order (with preseed)\n - All ISLs → all OSLs → all offsets\n - All ISLs → all OSLs → all offsets (same order, different algorithm)\n * - Top-up RNG\n - Continues from the same ``default_rng`` stream (``_preseed_rng``)\n - Continues from the same ``RandomState`` stream (``_preseed_rng``)\n\nBoth styles are preseeded whenever the corpus is\n:attr:`PromptCorpus.RANDOM` — the composer gates on\n``isinstance(dist, RangeRatioDistribution)``, which the SGLANG subclass\nsatisfies. ``_corpus_rng`` is the fallback for the non-preseeded paths\n(non-RANDOM corpora, and draws past cache exhaustion), not a per-style\ndifference.\n\n.. note::\n The two styles subtract special tokens at different points on purpose,\n each following its own upstream. VLLM folds the count into the bounds;\n SGLANG shifts each drawn length, mirroring ``sample_random_requests``\n under ``return_text`` (the default, and what aiperf sends)::\n\n input_lens[i] = max(1, input_lens[i] - num_special_tokens)\n\n Both land the same wire ISL on the configured value; only the resulting\n window shape differs.","enum":["vllm","sglang"],"title":"RandomCorpusStyle","type":"string"},"RankingsConfig":{"additionalProperties":false,"description":"Configuration for rankings/reranking endpoint datasets.\n\nControls the generation of query-passage pairs for benchmarking\nreranking and ranking models. Each request contains one query\nand multiple passages to rank.","properties":{"passages":{"description":"Number of passages per ranking request. Can be a fixed integer or {mean, stddev} distribution. Higher values test ranking at scale but increase request payload size and processing time.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Passages","x-jinja2-supported":true},"passageTokens":{"description":"Token length for each passage in ranking requests. Can be a fixed integer or {mean, stddev} distribution. Passages are synthetically generated text. Longer passages increase input processing demands and request size.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Passagetokens","x-jinja2-supported":true},"queryTokens":{"description":"Token length for the query text in ranking requests. Can be a fixed integer or {mean, stddev} distribution. Each ranking request contains one query and multiple passages.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Querytokens","x-jinja2-supported":true}},"title":"RankingsConfig","type":"object"},"RateSeriesConfig":{"additionalProperties":false,"description":"Piecewise-linear request-rate schedule.","properties":{"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"JSON file path containing request-rate control points.","title":"Path"},"points":{"description":"Strictly increasing request-rate control points.","items":{"$ref":"#/$defs/RateSeriesPoint"},"title":"Points","type":"array","minItems":2}},"title":"RateSeriesConfig","type":"object","allOf":[{"oneOf":[{"required":["path"],"properties":{"path":{"not":{"type":"null"}}}},{"required":["points"],"properties":{"points":{"minItems":2}}}]}]},"RateSeriesPoint":{"additionalProperties":false,"description":"One request-rate control point.","properties":{"timeS":{"description":"Elapsed time in seconds from the start of the request-rate series, after any request-rate ramp completes.","title":"Times","oneOf":[{"description":"Elapsed time in seconds from the start of the request-rate series, after any request-rate ramp completes.","minimum":0.0,"title":"Times","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"qps":{"description":"Request rate in queries per second at this point.","title":"Qps","oneOf":[{"description":"Request rate in queries per second at this point.","exclusiveMinimum":0.0,"title":"Qps","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["timeS","qps"],"title":"RateSeriesPoint","type":"object"},"RemoteFixedScheduleCheck":{"additionalProperties":false,"description":"Fixed-schedule auto-promotion deferred until a remote trace's first record is readable.\n\nSet by the CLI converter for a remote --input-file; consumed and cleared by the\ndataset access preflight.","properties":{"promoted":{"description":"True when the CLI built the profiling phase as fixed_schedule ahead of the timestamp check; the access preflight demotes it if the trace has none.","title":"Promoted","type":"boolean"},"conflictError":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Error raised if the trace turns out to have timestamps (an incompatible rate-shaped mode or a parameter sweep was selected).","title":"Conflicterror"},"demotedRequests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Request count the CLI would have defaulted to without the promotion (scaled to concurrency); applied if the access preflight demotes the phase. None when the user set a stop condition.","title":"Demotedrequests","x-jinja2-supported":true},"demotedStopWarning":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Warning logged when demoted_requests is applied.","title":"Demotedstopwarning"}},"required":["promoted"],"title":"RemoteFixedScheduleCheck","type":"object"},"RemoteRef":{"anyOf":[{"type":"string"},{"additionalProperties":{"type":"string"},"type":"object"}]},"RequestContentType":{"description":"Content type for HTTP request body serialization.","enum":["application/json","multipart/form-data"],"title":"RequestContentType","type":"string"},"RequestSignerType":{"enum":["sigv4"],"title":"RequestSignerType","type":"string"},"ResetKvCacheConfig":{"additionalProperties":false,"properties":{"timeoutSeconds":{"anyOf":[{"gt":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Timeout in seconds for the reset_kv_cache POST. When unset, uses 30.0s.","title":"Timeoutseconds","x-jinja2-supported":true},"maxRetrySeconds":{"anyOf":[{"ge":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total time budget in seconds for retrying a retryable reset_kv_cache POST failure: a timeout/connection error, or a response with status in RESET_KV_CACHE_RETRYABLE_STATUS_CODES (409/423/429/503). Other non-2xx responses are never retried. When unset, uses 60.0s.","title":"Maxretryseconds","x-jinja2-supported":true},"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for KV-cache reset. Default: /reset_prefix_cache.","title":"Path"}},"title":"ResetKvCacheConfig","type":"object"},"RuntimeConfig":{"additionalProperties":false,"description":"Runtime configuration for benchmark execution.","properties":{"ui":{"$ref":"#/$defs/UIType","default":"dashboard","description":"User interface mode. dashboard: rich interactive UI, simple: text progress, none: silent operation."},"workers":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total worker processes across the deployment (the maximum, ramped up to from `workers_min`). null = auto-detect based on CPU cores. Distinct from `workers_per_pod` (per-Kubernetes-pod fan-out) and `workers_min` (lower bound for adaptive ramp).","title":"Workers","x-jinja2-supported":true},"recordProcessors":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total parallel record processors across the deployment. null = auto-detect based on CPU cores. Distinct from `record_processors_per_pod` (per-Kubernetes-pod fan-out).","title":"Recordprocessors","x-jinja2-supported":true},"serviceRunType":{"$ref":"#/$defs/ServiceRunType","default":"multiprocessing","description":"Execution mode. multiprocessing: local multi-process (default for `aiperf profile`). kubernetes: distributed across pods. [operator-managed under AIPerfJob — the operator forces 'kubernetes' on the controller pod; do not set this in a CR spec.]"},"communication":{"anyOf":[{"discriminator":{"propertyName":"type","mapping":{"ipc":"#/$defs/IpcCommunicationConfig","tcp":"#/$defs/TcpCommunicationConfig","dual":"#/$defs/DualBindCommunicationConfig"}},"anyOf":[{"$ref":"#/$defs/IpcCommunicationConfig"},{"$ref":"#/$defs/TcpCommunicationConfig"},{"$ref":"#/$defs/DualBindCommunicationConfig"}]},{"type":"null"}],"default":null,"description":"Inter-process communication configuration. Defaults to IPC for single-machine operation.","title":"Communication"},"apiPort":{"anyOf":[{"maximum":65535,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AIPerf API server port. Enables HTTP and WebSocket endpoints for real-time metrics and control.","title":"Apiport","x-jinja2-supported":true},"apiHost":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"AIPerf API server host. Requires api_port to be set.","title":"Apihost"},"datasetApiBaseUrl":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Base URL the operator injects into worker pods so they can fetch datasets from the controller-pod sidecar in Kubernetes mode. [operator-managed; do not set in a CR spec.]","title":"Datasetapibaseurl"},"workersPerPod":{"anyOf":[{"maximum":100,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Worker containers packed into each Kubernetes worker pod (Kubernetes mode only). The total number of worker containers cluster-wide is `workers` (or auto-detected); this knob controls how that total is fanned across pods. Ignored in multiprocessing mode. Sizing guide: each worker was designed around ~500 concurrent connections, and a node's ~65k ephemeral ports is the ceiling this fan-out exists to work around.","title":"Workersperpod","x-jinja2-supported":true},"recordProcessorsPerPod":{"anyOf":[{"maximum":100,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Record-processor containers packed into each Kubernetes worker pod (Kubernetes mode only). Sibling of `workers_per_pod`; controls per-pod fan-out, not the cluster-wide total. Ignored in multiprocessing mode. Sizing guide: the design ratio is 1 record processor per 4 workers.","title":"Recordprocessorsperpod","x-jinja2-supported":true},"workersMin":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Lower bound on worker processes for the adaptive ramp. Distinct from `workers` (the upper bound / target total).","title":"Workersmin","x-jinja2-supported":true},"statsInterval":{"anyOf":[{"maximum":1000.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Interval in seconds between realtime stats publishes (dashboards and the per-tick log block). 0 disables the log block while dashboards continue to poll. Defaults to 5s under --ui dashboard, 30s otherwise. Overrides AIPERF_UI_REALTIME_METRICS_INTERVAL.","title":"Statsinterval","x-jinja2-supported":true}},"title":"RuntimeConfig","type":"object"},"SLAFilter":{"additionalProperties":false,"description":"SLA constraint applied to BO scoring or grid filtering.\n\nA trial is considered feasible iff ``stat(metric_tag) op threshold`` holds.\n``BayesianSearchPlanner`` consumes these for soft-penalty BO scoring and\n``write_search_history`` consumes them for lexicographic\nfeasibility-first best-result selection.\n\nLives in ``aiperf.config`` rather than ``aiperf.search_recipes`` because\n``AdaptiveSearchSweep`` carries a ``list[SLAFilter]`` field; the inverse\nlocation would force a circular import. Re-exported from\n``aiperf.search_recipes._base`` for ergonomic recipe authoring.\n\nExample: enforce p95 TTFT under 200 ms via\n``SLAFilter(metric_tag=\"time_to_first_token\", stat=\"p95\", op=\"lt\", threshold=200.0)``.","properties":{"metricTag":{"description":"Metric tag to filter on, e.g. 'time_to_first_token'. Must match a key in RunResult.summary_metrics produced by the run.","title":"Metrictag","type":"string"},"stat":{"default":"p95","description":"Statistic on the metric to compare against the threshold.","enum":["avg","p1","p5","p10","p25","p50","p75","p90","p95","p99","min","max"],"title":"Stat","type":"string"},"op":{"description":"Comparison operator. Filter passes when stat(metric) op threshold is true.","enum":["lt","le","gt","ge"],"title":"Op","type":"string"},"threshold":{"description":"Numeric threshold the metric statistic is compared against.","title":"Threshold","oneOf":[{"description":"Numeric threshold the metric statistic is compared against.","title":"Threshold","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metricTag","op","threshold"],"title":"SLAFilter","type":"object"},"SLOTier":{"additionalProperties":false,"description":"Named group of SLA filters representing one service-level objective.\n\nDefines a single tier for multi-tier SLO boundary search. Each tier\ncontains one or more SLA filters that must ALL pass for the tier to be\nconsidered feasible at a given concurrency level.\n\nLives in the config layer (alongside SLAFilter) to avoid circular imports\nbetween config and orchestrator packages. Re-exported from\n``aiperf.orchestrator.search_planner.multi_tier_models`` for ergonomic\norchestrator usage.","properties":{"label":{"description":"Unique tier identifier for output artifacts","title":"Label","type":"string"},"filters":{"description":"SLA filters that must ALL pass for this tier to be feasible","items":{"$ref":"#/$defs/SLAFilter"},"minItems":1,"title":"Filters","type":"array"},"orderingRank":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Rank in detected monotonic ordering (0=strictest); None if unordered","title":"Orderingrank","x-jinja2-supported":true}},"required":["label","filters"],"title":"SLOTier","type":"object"},"SageMakerConfig":{"additionalProperties":false,"description":"SageMaker Runtime routing options.\n\nGrouped rather than flattened onto ``EndpointConfig`` because they only\napply to one transport. Follows the same shape as ``reset_kv_cache`` /\n``server_profiler``.","properties":{"endpointName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Name of the SageMaker endpoint to invoke. Setting this is normally all that is needed: it selects the SageMaker transport, enables SigV4 signing, and derives the runtime base URL from aws_region.","title":"Endpointname"},"targetModel":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker TargetModel value, for multi-model endpoints only: SageMaker rejects requests that carry it on a single-model endpoint. Sent only when set -- it is a relative model-artifact path, so the request's model name is never used in its place. Not sent on streaming requests, which the API does not accept it on.","title":"Targetmodel"},"inferenceComponentName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker InferenceComponentName to target, for endpoints hosting multiple inference components.","title":"Inferencecomponentname"},"targetVariant":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker production variant to pin every request to, bypassing the endpoint's traffic split. Useful for benchmarking one variant of an A/B deployment in isolation.","title":"Targetvariant"}},"title":"SageMakerConfig","type":"object"},"SamplingDimension":{"additionalProperties":false,"description":"One dimension of a QMC sampling space.\n\nUsed by `SobolSweep` and `LatinHypercubeSweep`. Distinct from\n`SearchSpaceDimension` (BO) — same path/lo/hi grammar, but native\n`choices` support and explicit `scale` validation.\n\nPath is a dotted path of the form `phases.profiling.concurrency` —\nthe same grammar accepted by `_set_nested_value`.","properties":{"path":{"description":"Dotted-path into BenchmarkConfig.","title":"Path","type":"string"},"lo":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound (with hi).","title":"Lo","x-jinja2-supported":true},"hi":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound (with lo).","title":"Hi","x-jinja2-supported":true},"scale":{"default":"linear","description":"Mapping from unit cube to value.","enum":["linear","log"],"title":"Scale","type":"string"},"kind":{"default":"real","description":"Output type after mapping.","enum":["int","real"],"title":"Kind","type":"string"},"choices":{"anyOf":[{"items":{},"type":"array"},{"type":"null"}],"default":null,"description":"Discrete value list. Mutually exclusive with (lo, hi).","title":"Choices"}},"required":["path"],"title":"SamplingDimension","type":"object"},"ScenarioSweep":{"additionalProperties":false,"description":"Scenario sweep - hand-picked configurations deep-merged with base.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"scenarios","default":"scenarios","description":"Sweep type discriminator.","title":"Type","type":"string"},"runs":{"description":"List of scenario dicts to deep-merge with base config.","items":{"additionalProperties":true,"type":"object"},"minItems":1,"title":"Runs","type":"array"}},"required":["runs"],"title":"ScenarioSweep","type":"object"},"SearchPlannerType":{"enum":["bayesian","monotonic_sla","smooth_isotonic","optuna"],"title":"SearchPlannerType","type":"string"},"SearchSpaceDimension":{"additionalProperties":false,"description":"One dimension of the BO search space.\n\n`path` is a dotted path of the form `phases.profiling.concurrency` —\nthe same grammar accepted by `aiperf.config.sweep._set_nested_value`.","properties":{"path":{"description":"Dotted-path into BenchmarkConfig (e.g. 'phases.profiling.concurrency').","title":"Path","type":"string"},"lo":{"description":"Inclusive lower bound.","title":"Lo","oneOf":[{"description":"Inclusive lower bound.","title":"Lo","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"hi":{"description":"Inclusive upper bound.","title":"Hi","oneOf":[{"description":"Inclusive upper bound.","title":"Hi","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"kind":{"default":"real","description":"Dimension type. 'int' rounds planner suggestions to integers; 'real' keeps floats.","enum":["int","real"],"title":"Kind","type":"string"},"prior":{"default":"uniform","description":"Sampling prior for BO/Optuna planners. 'uniform' draws Sobol/random samples linearly across [lo, hi]; 'log-uniform' draws them log-spaced. Use 'log-uniform' when a parameter spans orders of magnitude (e.g. concurrency in [1, 1000]) so the initial-points phase covers low and high decades evenly. Requires lo > 0. Ignored by planners that do not draw initial points (monotonic, smooth_isotonic) where the search procedure is structurally adaptive rather than sampling-based.","enum":["uniform","log-uniform"],"title":"Prior","type":"string"}},"required":["path","lo","hi"],"title":"SearchSpaceDimension","type":"object"},"SequenceDistributionEntry":{"additionalProperties":false,"description":"Defines a single entry in an ISL/OSL probability distribution.\n\nAIPerf supports multi-modal token length distributions, allowing\ndifferent ISL/OSL combinations with relative frequencies for\nrealistic workload modeling.\n\nYAML Representation:\n sequence_distribution:\n - {isl: 128, osl: 64, probability: 40}\n - {isl: {mean: 512, stddev: 50}, osl: 256, probability: 35}\n - {isl: {mean: 2048, median: 1800}, osl: 512, probability: 25}","examples":[{"isl":128,"osl":64,"probability":40},{"isl":512,"isl_stddev":50,"osl":256,"osl_stddev":25,"probability":35}],"properties":{"isl":{"description":"Input sequence length (tokens). Can be a fixed integer, a {mean, stddev} distribution, or any distribution like {mean: 512, median: 400} for lognormal.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Isl","x-jinja2-supported":true},"islStddev":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Shorthand for ISL standard deviation. If provided when isl is an integer, creates a normal distribution. Cannot be used when isl is already a distribution dict.","title":"Islstddev","x-jinja2-supported":true},"osl":{"description":"Output sequence length (tokens). Can be a fixed integer, a {mean, stddev} distribution, or any typed distribution.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Osl","x-jinja2-supported":true},"oslStddev":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Shorthand for OSL standard deviation. If provided when osl is an integer, creates a normal distribution. Cannot be used when osl is already a distribution dict.","title":"Oslstddev","x-jinja2-supported":true},"probability":{"description":"Relative probability weight for this distribution bucket (0-100). Weights are normalized across all entries. Example: probability=40 means 40%% of requests use this ISL/OSL.","title":"Probability","oneOf":[{"description":"Relative probability weight for this distribution bucket (0-100). Weights are normalized across all entries. Example: probability=40 means 40%% of requests use this ISL/OSL.","maximum":100.0,"minimum":0.0,"title":"Probability","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["isl","osl","probability"],"title":"SequenceDistributionEntry","type":"object"},"ServerMetricsConfig":{"title":"ServerMetricsConfig","description":"Server metrics configuration for Prometheus scraping.\n\nCollects server-side operational metrics (queue depth, KV cache utilization,\nbatch sizes, GPU memory) from Prometheus endpoints exposed by inference servers\nlike vLLM, TensorRT-LLM, or Triton.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9090/metrics\"\n → ServerMetricsConfig(enabled=True, urls=[\"http://localhost:9090/metrics\"])\n - Singular url field: {url: \"...\"}\n → ServerMetricsConfig(urls=[\"...\"]) Can be a URL string or full config object.","anyOf":[{"additionalProperties":false,"description":"Server metrics configuration for Prometheus scraping.\n\nCollects server-side operational metrics (queue depth, KV cache utilization,\nbatch sizes, GPU memory) from Prometheus endpoints exposed by inference servers\nlike vLLM, TensorRT-LLM, or Triton.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9090/metrics\"\n → ServerMetricsConfig(enabled=True, urls=[\"http://localhost:9090/metrics\"])\n - Singular url field: {url: \"...\"}\n → ServerMetricsConfig(urls=[\"...\"])","properties":{"enabled":{"default":true,"description":"Enable Prometheus metrics scraping. Set to false to disable.","title":"Enabled","type":"boolean"},"urls":{"description":"Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","items":{"type":"string"},"title":"Urls","type":"array"},"formats":{"description":"Export formats for scraped metrics. Options: json, csv, parquet, jsonl.","items":{"$ref":"#/$defs/ServerMetricsFormat"},"title":"Formats","type":"array"},"discovery":{"$ref":"#/$defs/ServerMetricsDiscoveryConfig","description":"Auto-discovery of Prometheus endpoints in Kubernetes. Discovers pods via Dynamo labels, Prometheus annotations, or custom label selectors."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","oneOf":[{"type":"string"},{"description":"Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","items":{"type":"string"},"title":"Urls","type":"array"}],"x-singular-alias-of":"urls"}},"title":"ServerMetricsConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"string"}]},"ServerMetricsDiscoveryConfig":{"additionalProperties":false,"description":"Kubernetes-based auto-discovery of inference-server /metrics endpoints.\n\nWhen mode is 'auto' or 'kubernetes', queries the K8s API for pods that\nare recognizable inference servers (vLLM, SGLang, Triton Inference Server,\nTensorRT-LLM, NVIDIA Dynamo). Eligibility (any one is enough):\n1. Dynamo opt-in label: nvidia.com/metrics-enabled=true\n2. AIPerf opt-in annotation: aiperf.nvidia.com/metrics-paths=...\n3. A container image matching a known inference-server signature\n4. User-provided label_selector (server-side filter)\n\nThe broad ``prometheus.io/scrape=true`` annotation is intentionally NOT a\ntrigger: Loki, Grafana, kube-state-metrics, and many platform components\nset it without being inference servers. ``prometheus.io/{port,path,scheme}``\nare still honored to construct the scrape URL when an eligible pod sets them.","properties":{"mode":{"$ref":"#/$defs/ServerMetricsDiscoveryMode","default":"auto","description":"Discovery mode: 'auto' detects environment and tries K8s if in-cluster, 'kubernetes' forces K8s API discovery, 'disabled' uses only explicit URLs."},"labelSelector":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Kubernetes label selector for discovery. Example: 'app=vllm,env=prod'. Applied in addition to built-in Dynamo and Prometheus discovery.","title":"Labelselector"},"namespace":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Kubernetes namespace to search. If not specified, searches the benchmark pod's own namespace (the only namespace the default benchmark RBAC can list pods in). Use '*' to search all namespaces; this requires a cluster-scoped 'pods: list' grant on the benchmark ServiceAccount.","title":"Namespace"},"timeoutSeconds":{"default":30.0,"description":"Maximum seconds to wait for Kubernetes endpoint discovery before proceeding with explicit endpoints only.","title":"Timeoutseconds","oneOf":[{"default":30.0,"description":"Maximum seconds to wait for Kubernetes endpoint discovery before proceeding with explicit endpoints only.","exclusiveMinimum":0,"title":"Timeoutseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"ServerMetricsDiscoveryConfig","type":"object"},"ServerMetricsDiscoveryMode":{"description":"Mode for discovering server metrics endpoints in distributed environments.","enum":["auto","kubernetes","disabled"],"title":"ServerMetricsDiscoveryMode","type":"string"},"ServerMetricsFormat":{"description":"Format options for server metrics export.\n\nControls which output files are generated for server metrics data.\nDefault selection is JSON + CSV (JSONL excluded to avoid large files).","enum":["json","csv","jsonl","parquet"],"title":"ServerMetricsFormat","type":"string"},"ServerProfilerConfig":{"additionalProperties":false,"properties":{"urls":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"description":"Profiler control server URLs. When unset, uses the inference endpoint origins. Paths and queries are ignored.","title":"Urls"},"startBody":{"anyOf":[{"additionalProperties":{"$ref":"#/$defs/JsonValue"},"type":"object"},{"type":"null"}],"default":null,"description":"JSON object sent with profiler start requests. When unset, sends an empty body.","title":"Startbody","x-kubernetes-preserve-unknown-fields":true},"timeoutSeconds":{"anyOf":[{"gt":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Timeout in seconds for profiler start/stop POSTs. When unset, uses 30.0s.","title":"Timeoutseconds","x-jinja2-supported":true},"startPath":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for profiler start. Default: /start_profile.","title":"Startpath"},"stopPath":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for profiler stop. Default: /stop_profile.","title":"Stoppath"}},"title":"ServerProfilerConfig","type":"object"},"ServiceRunType":{"enum":["kubernetes","multiprocessing"],"title":"ServiceRunType","type":"string"},"SobolSweep":{"additionalProperties":false,"description":"Sobol quasi-Monte-Carlo sweep — low-discrepancy joint coverage.\n\nNote on ``scramble=False``: scipy's raw Sobol sequence starts at the\norigin (the first row is exactly all-zeros), so the first variant\nwill land on ``(lo, lo, ...)`` for every dimension. With small\n``samples`` this produces degenerate corner-only coverage. Prefer\nthe default ``scramble=True`` (Owen-scrambled) unless you have a\nspecific reason to want raw Sobol points.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"samples":{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","title":"Samples","oneOf":[{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","maximum":1048576,"minimum":2,"title":"Samples","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"seed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"RNG seed for reproducibility. Must be non-negative; scipy's QMC engines reject negative seeds with an opaque low-level error.","title":"Seed","x-jinja2-supported":true},"dimensions":{"description":"Dimensions to sample over.","items":{"$ref":"#/$defs/SamplingDimension"},"minItems":1,"title":"Dimensions","type":"array"},"labelFormat":{"default":"index","description":"Variation-label format. 'index' => '_NNNN' (short, sortable). 'kv' => 'k1=v1, k2=v2' (grid-style).","enum":["index","kv"],"title":"Labelformat","type":"string"},"type":{"const":"sobol","default":"sobol","description":"Sweep type discriminator.","title":"Type","type":"string"},"scramble":{"default":true,"description":"Owen-scrambled Sobol (recommended). False = raw Sobol sequence.","title":"Scramble","type":"boolean"}},"required":["samples","dimensions"],"title":"SobolSweep","type":"object"},"SweepMode":{"description":"Execution order for sweep + confidence composition.\n\nREPEATED: for trial in trials: for value in values\n All values tested per trial before moving to next trial.\n\nINDEPENDENT: for value in values: for trial in trials\n All trials completed per value before moving to next value.","enum":["independent","repeated"],"title":"SweepMode","type":"string"},"SynthesisConfig":{"additionalProperties":false,"description":"Configuration for trace synthesis/transformation.\n\nUsed with mooncake_trace format to transform production trace\ndata before replay. Allows scaling timestamps, token lengths,\nand radix tree structure.","properties":{"speedupRatio":{"default":1.0,"description":"Multiplier for timestamp scaling in synthesized traces. 1.0 = real-time, 2.0 = 2x faster, 0.5 = 2x slower.","title":"Speedupratio","oneOf":[{"default":1.0,"description":"Multiplier for timestamp scaling in synthesized traces. 1.0 = real-time, 2.0 = 2x faster, 0.5 = 2x slower.","exclusiveMinimum":0.0,"title":"Speedupratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"prefixLenMultiplier":{"default":1.0,"description":"Multiplier for core prefix branch lengths in the radix tree. 1.5 means prefix branches are 50%% longer.","title":"Prefixlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for core prefix branch lengths in the radix tree. 1.5 means prefix branches are 50%% longer.","exclusiveMinimum":0.0,"title":"Prefixlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"prefixRootMultiplier":{"default":1,"description":"Number of independent radix trees to distribute traces across. Higher values increase prefix diversity.","title":"Prefixrootmultiplier","oneOf":[{"default":1,"description":"Number of independent radix trees to distribute traces across. Higher values increase prefix diversity.","minimum":1,"title":"Prefixrootmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"promptLenMultiplier":{"default":1.0,"description":"Multiplier for leaf path (unique prompt) lengths. 2.0 means prompts are 2x longer.","title":"Promptlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for leaf path (unique prompt) lengths. 2.0 means prompts are 2x longer.","exclusiveMinimum":0.0,"title":"Promptlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"outputLenMultiplier":{"default":1.0,"description":"Multiplier for output lengths in synthesized traces.","title":"Outputlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for output lengths in synthesized traces.","minimum":0.0,"title":"Outputlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"maxIsl":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum input sequence length filter. Traces with input_length > max_isl are skipped entirely.","title":"Maxisl","x-jinja2-supported":true},"maxOsl":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum output sequence length cap. Traces with output_length > max_osl are capped to this value (not filtered).","title":"Maxosl","x-jinja2-supported":true},"allowDatasetWrap":{"default":false,"description":"Allow the finite weka/agentic trace pool to wrap (reuse distinct roots across concurrency lanes) when concurrency exceeds the number of eligible traces. Defaults to False: over-subscription fails unless this is explicitly enabled or an active cache-bust target already keeps repeated-trace traffic distinct.","title":"Allowdatasetwrap","type":"boolean"}},"title":"SynthesisConfig","type":"object"},"SyntheticDataset":{"additionalProperties":false,"description":"Synthetic dataset configuration.\n\nGenerates prompts programmatically based on token length\nspecifications. Ideal for controlled experiments.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"synthetic","description":"Dataset type discriminator. Must be 'synthetic'.","title":"Type","type":"string"},"entries":{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","title":"Entries","oneOf":[{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","minimum":1,"title":"Entries","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic dataset generation. When set, makes synthetic prompts, sampling, and other random operations reproducible across runs. Essential for A/B testing and debugging. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptConfig"},{"type":"null"}],"default":null,"description":"Prompt/token length configuration specifying ISL, OSL, sequence distributions, and batch processing settings."},"prefixPrompts":{"anyOf":[{"$ref":"#/$defs/PrefixPromptConfig"},{"type":"null"}],"default":null,"description":"Shared prefix configuration for KV cache testing. Generates prefix prompts that are prepended to user prompts, simulating cached context scenarios."},"turns":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Number of request-response turns per conversation. Can be a fixed integer or {mean, stddev} distribution. Each turn consists of a user message and model response. Set to 1 for single-turn interactions. Multi-turn conversations enable testing of context retention and conversation history handling.","title":"Turns"},"turnDelay":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Delay in milliseconds between consecutive turns within a multi-turn conversation. Can be a fixed value or {mean, stddev} distribution. Simulates user think time between receiving a response and sending the next message. Only applies when turns > 1. Set to 0 for back-to-back turns.","title":"Turndelay"},"turnDelayRatio":{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","title":"Turndelayratio","oneOf":[{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","minimum":0.0,"title":"Turndelayratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"images":{"anyOf":[{"$ref":"#/$defs/ImageConfig"},{"type":"null"}],"default":null,"description":"Synthetic image configuration for multimodal vision-language testing."},"audio":{"anyOf":[{"$ref":"#/$defs/AudioConfig"},{"type":"null"}],"default":null,"description":"Synthetic audio configuration for multimodal speech/audio testing."},"video":{"anyOf":[{"$ref":"#/$defs/VideoConfig"},{"type":"null"}],"default":null,"description":"Synthetic video configuration for multimodal video understanding testing."},"rankings":{"anyOf":[{"$ref":"#/$defs/RankingsConfig"},{"type":"null"}],"default":null,"description":"Rankings/reranking configuration for generating query-passage pairs. Only relevant for rankings endpoint types."},"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.isl; hoisted into prompts before validation.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.osl; hoisted into prompts before validation.","title":"Osl"}},"required":["name","type"],"title":"SyntheticDataset","type":"object"},"TcpCommunicationConfig":{"additionalProperties":false,"description":"TCP socket communication configuration.\n\nFor distributed deployments across machines.\nProvides detailed port configuration for all ZMQ proxies.","properties":{"type":{"const":"tcp","description":"Communication type. Must be 'tcp'.","title":"Type","type":"string"},"host":{"default":"127.0.0.1","description":"Host address for TCP communication. Use 0.0.0.0 to listen on all interfaces.","title":"Host","type":"string"},"recordsPort":{"default":5557,"description":"Port for records push/pull communication.","title":"Recordsport","oneOf":[{"default":5557,"description":"Port for records push/pull communication.","maximum":65535,"minimum":1,"title":"Recordsport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditRouterPort":{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","title":"Creditrouterport","oneOf":[{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Creditrouterport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"controlPort":{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","title":"Controlport","oneOf":[{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Controlport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditReturnPushPullPort":{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","title":"Creditreturnpushpullport","oneOf":[{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","maximum":65535,"minimum":1,"title":"Creditreturnpushpullport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"eventBusProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Event bus proxy ports (XPUB/XSUB)."},"datasetManagerProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Dataset manager proxy ports (DEALER/ROUTER)."},"rawInferenceProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Raw inference proxy ports (PUSH/PULL)."}},"required":["type"],"title":"TcpCommunicationConfig","type":"object"},"TcpProxyConfig":{"additionalProperties":false,"description":"TCP proxy port configuration for a single ZMQ proxy.","properties":{"frontendPort":{"description":"TCP port for proxy frontend.","title":"Frontendport","oneOf":[{"description":"TCP port for proxy frontend.","maximum":65535,"minimum":1,"title":"Frontendport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"backendPort":{"description":"TCP port for proxy backend.","title":"Backendport","oneOf":[{"description":"TCP port for proxy backend.","maximum":65535,"minimum":1,"title":"Backendport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["frontendPort","backendPort"],"title":"TcpProxyConfig","type":"object"},"TemplateConfig":{"additionalProperties":false,"description":"Configuration for custom template-based endpoints.\n\nWhen endpoint type is \"template\", this configures how requests\nare formatted and responses are parsed.","properties":{"body":{"description":"Jinja2 template string for request body. Variables: {{prompt}}, {{max_tokens}}, {{model}}, {{messages}}, etc.","title":"Body","type":"string"},"responseField":{"default":"text","description":"JSON path to extract response text from API response. Use dot notation for nested fields: 'choices.0.message.content'.","title":"Responsefield","type":"string"}},"required":["body"],"title":"TemplateConfig","type":"object"},"TimingMode":{"enum":["fixed_schedule","request_rate","adaptive_scale","user_centric_rate","agentic_replay"],"title":"TimingMode","type":"string"},"TokenizerConfig":{"additionalProperties":false,"description":"Tokenizer configuration for token counting and prompt generation.\n\nAIPerf uses a HuggingFace tokenizer for accurate token counting,\nwhich is essential for ISL/OSL enforcement and metrics calculation.","properties":{"name":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace tokenizer identifier, local filesystem path, or `builtin` for a zero-network-access tokenizer backed by tiktoken (o200k_base encoding). Should match the model's tokenizer for accurate token counts. If `--tokenizer` is not set and the model name looks like an obvious placeholder (e.g. `mock-model`, `test-model`, `fake-model`), AIPerf substitutes `builtin` automatically and emits a warning. Example: 'meta-llama/Llama-3.1-8B-Instruct'","title":"Name"},"revision":{"default":"main","description":"Model revision to use: branch name, tag, or commit hash. Use for version pinning to ensure reproducibility.","title":"Revision","type":"string"},"trustRemoteCode":{"default":false,"description":"Allow execution of custom tokenizer code from the repository. Required for some models but poses security risk. Only enable for trusted sources.","title":"Trustremotecode","type":"boolean"},"applyChatTemplate":{"default":false,"description":"Apply the HuggingFace tokenizer's chat template when counting input tokens. When enabled: synthetic ISL is compensated for chat-template wrapping (BOS, role headers, EOT, generation-prompt suffix) and the record processor reports ISL using `apply_chat_template(tokenize=True, add_generation_prompt=True)` for chat-shape payloads. When disabled (default), both paths use bare-text encoding, so reported ISL matches the prompt content the user asked for and ignores template overhead. Requires an HF tokenizer with a chat template configured; no-ops on tiktoken / un-templated models.","title":"Applychattemplate","type":"boolean"}},"title":"TokenizerConfig","type":"object"},"TokenizerOverride":{"additionalProperties":false,"description":"Per-model tokenizer override configuration.\n\nAllows specifying a different tokenizer for a specific model,\nuseful when models require specialized tokenization.","properties":{"name":{"description":"HuggingFace tokenizer identifier or local filesystem path.","title":"Name","type":"string"}},"required":["name"],"title":"TokenizerOverride","type":"object"},"TransportType":{"enum":["http","sagemaker"],"title":"TransportType","type":"string"},"UIType":{"enum":["dashboard","none","simple"],"title":"UIType","type":"string"},"URLSelectionStrategy":{"enum":["round_robin"],"title":"URLSelectionStrategy","type":"string"},"UserCentricPhase":{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","title":"Users","oneOf":[{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},"UserFile":{"description":"One user-declared output file rendered into the run directory before benchmark start.\n\nPath is relative to the run directory; subdirectories are allowed; absolute\npaths and any segment equal to '..' are rejected. Content is rendered with\njinja2 against a documented context (variables: + system-injected names).","properties":{"path":{"description":"Output path relative to the run directory. Subdirectories allowed. Absolute paths and any segment equal to '..' are rejected.","title":"Path","type":"string"},"format":{"anyOf":[{"enum":["json","yaml","text"],"type":"string"},{"type":"null"}],"default":null,"description":"Serialization format. If omitted: 'text' when content is a string, 'json' otherwise.","title":"Format"},"content":{"description":"Templated value. Structured (dict/list/scalar) for json/yaml; string for text. Jinja2 expressions in any string leaf are rendered with the user_files context.","title":"Content","x-kubernetes-preserve-unknown-fields":true}},"required":["path","content"],"title":"UserFile","type":"object"},"VideoAudioCodec":{"description":"Audio codecs for embedding audio in synthetic video files.","enum":["aac","libvorbis","libopus"],"title":"VideoAudioCodec","type":"string"},"VideoAudioConfig":{"additionalProperties":false,"description":"Configuration for embedding an audio track in synthetic video files.","properties":{"sampleRate":{"default":44.1,"description":"Audio sample rate in kHz for the embedded audio track. Common values: 8 (telephony), 16 (speech), 44.1 (CD quality), 48 (professional). Higher sample rates increase audio fidelity and file size.","title":"Samplerate","oneOf":[{"default":44.1,"description":"Audio sample rate in kHz for the embedded audio track. Common values: 8 (telephony), 16 (speech), 44.1 (CD quality), 48 (professional). Higher sample rates increase audio fidelity and file size.","maximum":96.0,"minimum":8.0,"title":"Samplerate","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"channels":{"default":0,"description":"Number of audio channels to embed in generated video files. 0 = disabled (no audio track, default), 1 = mono, 2 = stereo. When set to 1 or 2, a Gaussian noise audio track matching the video duration is muxed into each video via FFmpeg.","title":"Channels","oneOf":[{"default":0,"description":"Number of audio channels to embed in generated video files. 0 = disabled (no audio track, default), 1 = mono, 2 = stereo. When set to 1 or 2, a Gaussian noise audio track matching the video duration is muxed into each video via FFmpeg.","maximum":2,"minimum":0,"title":"Channels","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"codec":{"anyOf":[{"$ref":"#/$defs/VideoAudioCodec"},{"type":"null"}],"default":null,"description":"Audio codec for the embedded audio track. If not specified, auto-selects based on video format: libopus for MP4, libvorbis for WebM. Options: libvorbis, libopus, aac. The AIPerf container ships only libvorbis and libopus; aac requires an FFmpeg build that includes an AAC encoder. libopus always encodes at 48 kHz, so any --video-audio-sample-rate is resampled during muxing."},"depth":{"description":"Audio bit depth for the embedded audio track. Supported values: 8, 16, 24, or 32 bits. Higher bit depths provide greater dynamic range but increase file size.","default":16,"oneOf":[{"default":16,"description":"Audio bit depth for the embedded audio track. Supported values: 8, 16, 24, or 32 bits. Higher bit depths provide greater dynamic range but increase file size.","enum":[8,16,24,32],"title":"Depth","type":"integer"},{"type":"string","enum":["8","16","24","32"]},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"VideoAudioConfig","type":"object"},"VideoConfig":{"additionalProperties":false,"description":"Configuration for synthetic video generation in multimodal datasets.\n\nControls the generation of synthetic videos for video understanding\nmodel benchmarking. Requires FFmpeg for video generation.","properties":{"batchSize":{"default":0,"description":"Number of video files to include in each multimodal request. Supported with chat endpoint type for video understanding models. Set to 0 to disable video inputs. Higher batch sizes significantly increase request payload size.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of video files to include in each multimodal request. Supported with chat endpoint type for video understanding models. Set to 0 to disable video inputs. Higher batch sizes significantly increase request payload size.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"duration":{"default":1.0,"description":"Duration in seconds for each generated video clip. Combined with fps, determines total frame count (frames = duration * fps). Longer durations increase file size and processing time. Typical values: 1-10 seconds for testing.","title":"Duration","oneOf":[{"default":1.0,"description":"Duration in seconds for each generated video clip. Combined with fps, determines total frame count (frames = duration * fps). Longer durations increase file size and processing time. Typical values: 1-10 seconds for testing.","exclusiveMinimum":0.0,"title":"Duration","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"fps":{"default":4,"description":"Frames per second for generated video. Higher FPS creates smoother video but increases frame count and file size. Common values: 4 (minimal, recommended for Cosmos models), 24 (cinematic), 30 (standard), 60 (high frame rate). Total frames = duration * fps.","title":"Fps","oneOf":[{"default":4,"description":"Frames per second for generated video. Higher FPS creates smoother video but increases frame count and file size. Common values: 4 (minimal, recommended for Cosmos models), 24 (cinematic), 30 (standard), 60 (high frame rate). Total frames = duration * fps.","minimum":1,"title":"Fps","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"width":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Video frame width in pixels. Determines video resolution and file size. Common values: 640 (SD), 1280 (HD), 1920 (Full HD). If not specified, uses codec/format defaults.","title":"Width","x-jinja2-supported":true},"height":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Video frame height in pixels. Combined with width determines aspect ratio and total pixel count per frame. Common values: 480 (SD), 720 (HD), 1080 (Full HD). If not specified, uses codec/format defaults.","title":"Height","x-jinja2-supported":true},"format":{"$ref":"#/$defs/VideoFormat","default":"webm","description":"Container format for generated video files. webm: VP9 codec, BSD-licensed, recommended for open-source workflows. mp4: widely compatible container, VP9 by default. Format affects compatibility, file size, and encoding options."},"codec":{"default":"libvpx-vp9","description":"Video codec for encoding. Common options: libvpx-vp9 (CPU, BSD-licensed, default), libvpx (CPU, BSD-licensed, VP8). Any codec the local FFmpeg supports can be used, but the AIPerf container ships a minimal FFmpeg build limited to VP8/VP9 video with Vorbis/Opus audio; codecs such as libx264 or h264_nvenc require an FFmpeg build that includes them.","title":"Codec","type":"string"},"synthType":{"$ref":"#/$defs/VideoSynthType","default":"moving_shapes","description":"Algorithm for generating synthetic video content. Different types produce different visual patterns for testing. Content doesn't affect semantic meaning but may impact encoding efficiency and file size."},"audio":{"$ref":"#/$defs/VideoAudioConfig","description":"Audio track configuration for embedding audio in generated videos."}},"title":"VideoConfig","type":"object"},"VideoFormat":{"description":"Video container formats for synthetic video generation.","enum":["mp4","webm"],"title":"VideoFormat","type":"string"},"VideoSynthType":{"enum":["moving_shapes","grid_clock","noise"],"title":"VideoSynthType","type":"string"},"WandbConfig":{"additionalProperties":false,"dependentRequired":{"entity":["project"],"runName":["project"],"tags":["project"]},"description":"Weights & Biases run-upload configuration.","properties":{"project":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases project name. Setting this enables wandb export.","title":"Project"},"entity":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases entity (team or user). Defaults to the API key's default entity.","title":"Entity"},"runName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases run name.","title":"Runname"},"tags":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Additional Weights & Biases run tags to attach on upload.","title":"Tags"}},"title":"WandbConfig","type":"object"},"ZipSweep":{"additionalProperties":false,"description":"Zip sweep — parameters paired element-wise (lockstep).\n\nLike ``GridSweep`` but uses ``zip(strict=True)`` instead of Cartesian\nproduct. All parameter lists must have identical length. Use when you\nwant N runs each setting a coordinated tuple of fields, without the\nNxM blow-up of grid. Canonical use case: paired ISL/OSL.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"zip","default":"zip","description":"Sweep type discriminator.","title":"Type","type":"string"},"parameters":{"additionalProperties":{"items":{},"type":"array"},"description":"Parameters to sweep in lockstep: dot-notation path -> list of values. All lists must have equal length.","minProperties":1,"title":"Parameters","type":"object"}},"required":["parameters"],"title":"ZipSweep","type":"object"},"AdaptiveScaleNestedBlock":{"type":"object","additionalProperties":false,"properties":{"enabled":{"anyOf":[{"type":"boolean"},{"type":"string","enum":["true","false","yes","no","on","off","1","0"]}]},"control":{"type":"object","additionalProperties":false,"properties":{"variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"strategy":{"type":"object","additionalProperties":false,"properties":{"type":{"type":"string","enum":["ramp_until_fail"]},"stepPolicy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"step_policy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"baseStep":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"base_step":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxStepMultiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_step_multiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"stepPercent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"step_percent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"sla":{"anyOf":[{"type":"array","items":{"$ref":"#/$defs/SLAFilter"}},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"controlVariable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"control_variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"controlMin":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"controlMax":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"window":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessmentPeriod":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessment_period":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minCompletedRequests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_completed_requests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustainDuration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustain_duration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"DisabledAdaptiveScaleNestedBlock":{"type":"object","additionalProperties":false,"properties":{"enabled":{"anyOf":[{"const":false},{"type":"string","enum":["false","no","off","0"]}]},"control":{"type":"object","additionalProperties":false,"properties":{"variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"strategy":{"type":"object","additionalProperties":false,"properties":{"type":{"type":"string","enum":["ramp_until_fail"]},"stepPolicy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"step_policy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"baseStep":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"base_step":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxStepMultiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_step_multiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"stepPercent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"step_percent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"sla":{"anyOf":[{"type":"array","items":{"$ref":"#/$defs/SLAFilter"}},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"controlVariable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"control_variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"controlMin":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"controlMax":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"window":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessmentPeriod":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessment_period":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minCompletedRequests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_completed_requests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustainDuration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustain_duration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}},"required":["enabled"]}},"additionalProperties":false,"description":"JSON Schema for AIPerf YAML configuration files. This schema enables IDE autocompletion and validation for AIPerf benchmark configurations. Schema version: 2.0.0","properties":{"schemaVersion":{"const":"2.0","default":"2.0","description":"AIPerf config schema version. Pinned to '2.0' for the envelope/body shape (envelope keys at top level, swept body under ``benchmark``).","title":"Schemaversion","type":"string"},"benchmark":{"$ref":"#/$defs/BenchmarkConfig","description":"Benchmark workload (the swept body)."},"sweep":{"anyOf":[{"discriminator":{"propertyName":"type","mapping":{"grid":"#/$defs/GridSweep","zip":"#/$defs/ZipSweep","scenarios":"#/$defs/ScenarioSweep","adaptive_search":"#/$defs/AdaptiveSearchSweep","sobol":"#/$defs/SobolSweep","latin_hypercube":"#/$defs/LatinHypercubeSweep"}},"anyOf":[{"$ref":"#/$defs/GridSweep"},{"$ref":"#/$defs/ZipSweep"},{"$ref":"#/$defs/ScenarioSweep"},{"$ref":"#/$defs/AdaptiveSearchSweep"},{"$ref":"#/$defs/SobolSweep"},{"$ref":"#/$defs/LatinHypercubeSweep"}]},{"type":"null"}],"default":null,"description":"Sweep configuration for parameter exploration. Supports grid (Cartesian product), zip (lock-stepped axes), scenarios (hand-picked), sobol / latin_hypercube (QMC sampling), and adaptive_search (Bayesian / monotonic outer-loop) strategies.","title":"Sweep"},"multiRun":{"$ref":"#/$defs/MultiRunConfig","description":"Multi-run benchmarking configuration for statistical reporting. When num_runs > 1, executes multiple runs and computes aggregate statistics."},"plot":{"default":null,"description":"Plot configuration for this run/sweep. When set, ``~/.aiperf/plot_config.yaml`` is ignored. May be a bare-string path to a plot YAML (resolved relative to the AIPerf config file's directory, or absolute), or an inline mapping mirroring ``default_plot_config.yaml``. Presence implies ``artifacts.auto_plot=True`` unless explicitly set False.","title":"Plot"},"variables":{"additionalProperties":true,"description":"User-defined values exposed to Jinja2 in `{{ ... }}` expressions during config load. Cross-variation: scenario `runs[i].variables:` deep-merge over this base. Preserved on the resolved config so run-time renderers can resolve them again.","title":"Variables","type":"object"},"randomSeed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Global random seed for reproducibility. Base seed for per-variation derivation in sweep mode (variation N gets base + N). Must be non-negative; numpy and scipy reject negative seeds.","title":"Randomseed","x-jinja2-supported":true},"noSweepTable":{"default":false,"description":"Suppress the per-cell streaming sweep table during multi-variation sweeps. Auto-suppressed when stdout is non-interactive, when the dashboard UI is active, or for single-cell sweeps.","title":"Nosweeptable","type":"boolean"}},"required":["benchmark"],"title":"AIPerf Configuration Schema","type":"object","$comment":"Auto-generated by tools/generate_config_schema.py. Do not edit manually.","x-taplo":{"initKeys":["benchmark","sweep","multiRun","variables","randomSeed"]}} +{"$schema":"https://json-schema.org/draft/2020-12/schema","$id":"https://nvidia.github.io/aiperf/schemas/aiperf-config.schema.json","$defs":{"AIPerfLogLevel":{"description":"Logging levels for AIPerf output verbosity.","enum":["TRACE","DEBUG","INFO","NOTICE","WARNING","SUCCESS","ERROR","CRITICAL"],"title":"AIPerfLogLevel","type":"string"},"AccuracyBenchmarkType":{"enum":["mmlu","mmlu_pro","aime","hellaswag","bigbench","aime24","aime25","math_500","gsm8k","gpqa_diamond","lcb_codegeneration"],"title":"AccuracyBenchmarkType","type":"string"},"AccuracyConfig":{"additionalProperties":false,"description":"Configuration for accuracy benchmarking mode.\n\nWhen benchmark is set, enables accuracy evaluation alongside\nperformance profiling using standard benchmarks (MMLU, AIME, etc.).","properties":{"benchmark":{"anyOf":[{"$ref":"#/$defs/AccuracyBenchmarkType"},{"type":"null"}],"default":null,"description":"Accuracy benchmark to run. When set, enables accuracy benchmarking alongside performance profiling. AIME variants: 'aime' is the legacy combined set (deprecated for new runs); prefer the year-pinned 'aime24' or 'aime25' for reproducibility."},"tasks":{"anyOf":[{"type":"string"},{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Specific tasks or subtasks within the benchmark to evaluate (e.g., specific MMLU subjects). Accepts comma-separated values (e.g. abstract_algebra,anatomy) or repeated flags. If not set, all tasks are included.","title":"Tasks"},"nShots":{"anyOf":[{"maximum":32,"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of few-shot examples to include in the prompt. 0 means zero-shot evaluation, None uses the benchmark default (e.g. MMLU=5). Maximum 32.","title":"Nshots","x-jinja2-supported":true},"enableCot":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Enable chain-of-thought prompting for accuracy evaluation. Adds reasoning instructions to the prompt. Defaults to the benchmark's ``default_enable_cot`` metadata when unset (e.g. AIME defaults to True).","title":"Enablecot"},"grader":{"anyOf":[{"$ref":"#/$defs/AccuracyGraderType"},{"type":"null"}],"default":null,"description":"Override the default grader for the selected benchmark (e.g., exact_match, math, multiple_choice, code_execution). If not set, uses the benchmark's default grader."},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Custom system prompt to use for accuracy evaluation. Overrides any benchmark-specific system prompt.","title":"Systemprompt"},"verbose":{"default":false,"description":"Enable verbose output for accuracy evaluation, showing per-problem grading details.","title":"Verbose","type":"boolean"}},"title":"AccuracyConfig","type":"object"},"AccuracyGraderType":{"enum":["exact_match","math","multiple_choice","code_execution","lighteval_expr","lighteval_latex","lighteval_gpqa","lighteval_gsm8k","mmlu_pro"],"title":"AccuracyGraderType","type":"string"},"AdaptiveSearchSweep":{"additionalProperties":false,"description":"Adaptive outer-loop search (Bayesian / monotonic / etc.).\n\nA third sweep variant: BO drives variation choice itself, so the\ngrid-style ``iteration_order`` and ``same_seed`` knobs do not apply.\nInherits cooldown/sla_filters/post_process from `_SweepBase` only.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"type":{"const":"adaptive_search","default":"adaptive_search","description":"Sweep type discriminator.","title":"Type","type":"string"},"planner":{"$ref":"#/$defs/SearchPlannerType","default":"bayesian","description":"Outer-loop planner plugin name (bayesian | optuna | monotonic_sla | smooth_isotonic)."},"searchSpace":{"description":"Dimensions to optimize over.","items":{"$ref":"#/$defs/SearchSpaceDimension"},"minItems":1,"title":"Searchspace","type":"array"},"objectives":{"description":"What to optimize. Length-1 = single-objective search; length-N = multi-objective search. BoTorch-backed Pareto BO requires --optuna-sampler botorch with a multi-objective acquisition like qlognehvi.","items":{"$ref":"#/$defs/Objective"},"minItems":1,"title":"Objectives","type":"array"},"outcomeConstraints":{"description":"Feasibility gates on non-optimized metrics. Empty = no constraints.","items":{"$ref":"#/$defs/OutcomeConstraint"},"title":"Outcomeconstraints","type":"array"},"maxIterations":{"description":"Maximum BO iterations.","title":"Maxiterations","oneOf":[{"description":"Maximum BO iterations.","maximum":200,"minimum":2,"title":"Maxiterations","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"nInitialPoints":{"default":5,"description":"Sobol-random points before the GP fits. Must be < max_iterations for `bayesian`/`optuna` planners; ignored by 1-D SLA planners.","title":"Ninitialpoints","oneOf":[{"default":5,"description":"Sobol-random points before the GP fits. Must be < max_iterations for `bayesian`/`optuna` planners; ignored by 1-D SLA planners.","minimum":1,"title":"Ninitialpoints","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"plateauWindow":{"default":8,"description":"Recent-iteration window for plateau detection. Default 8 keeps the window strictly larger than ``n_initial_points`` (default 5), so the plateau test always inspects at least a few GP-driven iterations rather than tripping on the Sobol-only prefix when the initial points happen to land in a flat region.","title":"Plateauwindow","oneOf":[{"default":8,"description":"Recent-iteration window for plateau detection. Default 8 keeps the window strictly larger than ``n_initial_points`` (default 5), so the plateau test always inspects at least a few GP-driven iterations rather than tripping on the Sobol-only prefix when the initial points happen to land in a flat region.","minimum":2,"title":"Plateauwindow","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"plateauThreshold":{"default":0.01,"description":"Coefficient-of-variation threshold for plateau stop.","title":"Plateauthreshold","oneOf":[{"default":0.01,"description":"Coefficient-of-variation threshold for plateau stop.","exclusiveMinimum":0,"title":"Plateauthreshold","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"improvementPatience":{"default":10,"description":"Stop after this many no-improvement iterations.","title":"Improvementpatience","oneOf":[{"default":10,"description":"Stop after this many no-improvement iterations.","minimum":2,"title":"Improvementpatience","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"randomSeed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"If set, passed as `random_state` to the planner backend.","title":"Randomseed","x-jinja2-supported":true},"optunaSampler":{"default":"botorch","description":"Optuna sampler when planner=optuna. Default ``botorch`` is the preferred Gaussian-process qLogNEI/qLogNEHVI path and requires the optional ``botorch`` extra; when this default is implicit and the optional stack is unavailable, the planner warns and falls back to ``tpe``. ``tpe`` is the dep-light Tree-Parzen estimator that ships with Optuna core. ``gp`` is Optuna's pure-GP sampler and requires torch.","enum":["gp","tpe","botorch"],"title":"Optunasampler","type":"string"},"optunaAcquisition":{"anyOf":[{"enum":["logei","qlogei","qnei","qlognei","qehvi","qnehvi","qlognehvi"],"type":"string"},{"type":"null"}],"default":null,"description":"Acquisition override for the BoTorch sampler. None lets Optuna pick (single-objective LogEI). ``qnei`` = Letham 2017 noisy-EI; ``qlognei`` = Ament 2023 (BoTorch's strongly recommended modern default; requires botorch>=0.10). ``qehvi``/``qnehvi``/``qlognehvi`` are multi-objective hypervolume acquisitions for Pareto BO. Only meaningful when planner=optuna and optuna_sampler=botorch; ignored otherwise.","title":"Optunaacquisition"},"optunaTerminator":{"default":"none","description":"Optional posterior-regret stopping rule (Makarova 2022 / Ishibashi 2023; same family as Wilson 2024). Layered on top of three-signal convergence. Only meaningful when planner=optuna; ignored otherwise.","enum":["regret","emmr","none"],"title":"Optunaterminator","type":"string"},"objectivePooling":{"default":"mean","description":"Per-trial percentile aggregation. ``mean`` = arithmetic mean of per-trial percentiles across trials (current default). ``pooled`` = walks each trial's profile_export.jsonl and computes np.percentile over the pooled raw-sample bag (the statistic that satisfies SLO claims; requires --export-level records). Only meaningful when objective.stat is a percentile (p50/p90/p95/p99); a no-op when stat=avg.","enum":["mean","pooled"],"title":"Objectivepooling","type":"string"},"monotonicStabilityTrials":{"default":2,"description":"MonotonicSLA planner stability-window size.","title":"Monotonicstabilitytrials","oneOf":[{"default":2,"description":"MonotonicSLA planner stability-window size.","minimum":1,"title":"Monotonicstabilitytrials","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaReplicates":{"default":0,"description":"Replicate-step trial count override for the smooth_isotonic SLA planner. 0 = auto (Hyperband-flavored formula based on sigma_margin/threshold). >0 = fixed override. Ignored by other planners.","title":"Slareplicates","oneOf":[{"default":0,"description":"Replicate-step trial count override for the smooth_isotonic SLA planner. 0 = auto (Hyperband-flavored formula based on sigma_margin/threshold). >0 = fixed override. Ignored by other planners.","minimum":0,"title":"Slareplicates","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaPrecision":{"default":"normal","description":"Per-probe sample-budget knob for the smooth_isotonic SLA planner. tight=10000, normal=1000, coarse=300 requests. Drives p99 CI width. Ignored by other planners.","enum":["tight","normal","coarse"],"title":"Slaprecision","type":"string"},"slaWarmupSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Discard this many seconds at the start of each probe before computing margins, to skip CUDA-graph warmup and KV-cache shape transients. None = auto (max(30s, 3 * inter-batch-time)). Ignored by other planners.","title":"Slawarmupseconds","x-jinja2-supported":true},"constraintMode":{"default":"penalty","description":"Constrained-BO acquisition mode (penalty | eic).","enum":["penalty","eic"],"title":"Constraintmode","type":"string"},"slaTiers":{"description":"Multi-tier SLO definitions parsed from --search-sla-tier flags. Each tier is a named group of SLA filters sharing a common boundary search. When non-empty (2-10 tiers), the MultiTierPlanner is activated instead of single-tier planners.","items":{"$ref":"#/$defs/SLOTier"},"title":"Slatiers","type":"array"}},"required":["searchSpace","objectives","maxIterations"],"title":"AdaptiveSearchSweep","type":"object"},"ArtifactsConfig":{"additionalProperties":false,"description":"Artifacts configuration for benchmark output.\n\nControls where and how benchmark results are exported.\nUses flat structure with format arrays instead of nested export configs.","properties":{"dir":{"default":"artifacts","description":"Output directory for all benchmark artifacts. Created if it doesn't exist.","format":"path","title":"Dir","type":"string"},"prefix":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Base filename override applied to ALL profile and server-metrics exports. With prefix='foo' every output becomes `foo.csv`, `foo.json`, `foo_timeslices.{csv,json}`, `foo.jsonl`, `foo_raw.jsonl`, `foo_outputs.json`, `foo_gpu_telemetry.jsonl`, `foo_server_metrics.{jsonl,json,csv,parquet}`. When unset (the default), historical per-file names are used: `profile_export_aiperf.csv/json`, `profile_export.jsonl`, `profile_export_raw.jsonl`, `outputs.json`, `gpu_telemetry_export.jsonl`, `server_metrics_export.{jsonl,json,csv,parquet}`. Known suffixes (`_raw.jsonl`, `_outputs.json`, `_timeslices.{csv,json}`, `_gpu_telemetry.jsonl`, `_server_metrics.{jsonl,json,csv,parquet}`, `.csv`/`.json`/`.jsonl`/`.parquet`) are stripped from the supplied value so `--profile-export-prefix foo_raw.jsonl` still yields a clean `foo` base.","title":"Prefix"},"summary":{"anyOf":[{"items":{"const":"json","type":"string"},"type":"array"},{"const":false,"type":"boolean"}],"description":"Summary export formats. Only 'json' is wired up to this field; the CSV summary is emitted regardless. Set to false to disable the summary JSON file only.","title":"Summary"},"records":{"anyOf":[{"items":{"const":"jsonl","type":"string"},"type":"array"},{"const":false,"type":"boolean"}],"description":"Per-request records export formats. Only 'jsonl' is wired up today. Set to false to disable the per-record JSONL file.","title":"Records"},"raw":{"default":false,"description":"Export raw request/response payloads as JSONL.","title":"Raw","type":"boolean"},"trace":{"default":false,"description":"Export HTTP trace data for debugging.","title":"Trace","type":"boolean"},"sliceDuration":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Time slice duration in seconds for trend analysis (must be > 0). Divides benchmark into windows for per-window statistics. Supports: 30, '30s', '5m', '2h'.","title":"Sliceduration","x-jinja2-supported":true},"showTraceTiming":{"default":false,"description":"Display HTTP trace timing metrics in console output. Shows detailed timing breakdown: blocked, DNS, connecting, sending, waiting (TTFB), receiving, and total duration.","title":"Showtracetiming","type":"boolean"},"userFiles":{"description":"User-defined templated files materialized into the run directory before the benchmark begins.","items":{"$ref":"#/$defs/UserFile"},"title":"Userfiles","type":"array"},"autoPlot":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Auto-invoke `aiperf plot` against the artifact directory after the benchmark completes. Unset (the default) means a `plot:` section implies True; set False to suppress plots even with one, True to force them. `AIPerfConfig` resolves this to a concrete bool, so any config that came through the loader carries a plain bool.","title":"Autoplot"},"plotRequired":{"default":false,"description":"Treat auto-plot failures as fatal: re-raise so `aiperf profile` exits non-zero. Only meaningful when auto_plot is True. Default False = warn and continue.","title":"Plotrequired","type":"boolean"},"exportOutputsJson":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"Export generated response text after the run, to `outputs.json` or `_outputs.json` when `prefix` is set. Unset (the default) means the `raw` export level implies it; set False to opt out at raw, True to force it at any level. Always a concrete bool once validation has run.","title":"Exportoutputsjson"}},"title":"ArtifactsConfig","type":"object"},"AudioConfig":{"additionalProperties":false,"description":"Configuration for synthetic audio generation in multimodal datasets.\n\nControls the generation of synthetic audio for speech-to-text\nand audio-language model benchmarking. Generated audio is random\nnoise with specified sample rate, bit depth, and format.","properties":{"batchSize":{"default":0,"description":"Number of audio inputs to include in each multimodal request. Supported with chat endpoint type for multimodal models. Set to 0 to disable audio inputs.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of audio inputs to include in each multimodal request. Supported with chat endpoint type for multimodal models. Set to 0 to disable audio inputs.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"length":{"description":"Audio duration in seconds. Can be a fixed value or {mean, stddev} distribution. Used when batch_size > 0 for multimodal benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Length","x-jinja2-supported":true},"format":{"$ref":"#/$defs/AudioFormat","default":"wav","description":"File format for generated audio files. wav: uncompressed PCM (larger files). mp3: compressed (smaller files). Format affects file size in multimodal requests but not audio characteristics."},"sampleRates":{"description":"List of audio sample rates in kHz to randomly select from. Common values: 8.0 (telephony), 16.0 (speech), 44.1 (CD quality), 48.0 (professional). Specify multiple values for mixed-quality testing.","items":{"type":"number"},"minItems":1,"title":"Samplerates","type":"array"},"depths":{"description":"List of audio bit depths in bits to randomly select from. Each audio file is assigned a random depth from this list. Common values: 8 (low quality), 16 (CD quality), 24 (professional), 32 (high-end). Specify multiple values for mixed-quality testing.","items":{"type":"integer"},"minItems":1,"title":"Depths","type":"array"},"channels":{"default":1,"description":"Number of audio channels. 1 = mono (single channel), 2 = stereo (left/right channels). Stereo doubles file size. Most speech models use mono.","title":"Channels","oneOf":[{"default":1,"description":"Number of audio channels. 1 = mono (single channel), 2 = stereo (left/right channels). Stereo doubles file size. Most speech models use mono.","maximum":2,"minimum":1,"title":"Channels","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"AudioConfig","type":"object"},"AudioFormat":{"description":"Audio file formats for synthetic audio generation.","enum":["wav","mp3"],"title":"AudioFormat","type":"string"},"BenchmarkConfig":{"additionalProperties":false,"description":"Pure runtime configuration - what SystemController and services need.\n\nContains all fields required to execute a single benchmark run.\nDoes NOT include sweep, multi_run, variables, or random_seed settings\n(those live on AIPerfConfig as envelope-level cross-variation fields).\n\nRequired Sections:\n - models: Model(s) to benchmark\n - endpoint: Server connection settings\n - datasets: Single-element list of dataset configuration\n - phases: Ordered list of benchmark phases. A single-dict under\n ``phases:`` is rejected; use the envelope-level ``warmup:`` /\n ``profiling:`` shorthand or an explicit list.\n\nOptional Sections:\n - artifacts: Export and console settings\n - slos: SLO-based quality metrics (generic dict)\n - tokenizer: Token counting configuration\n - gpu_telemetry: GPU metrics from DCGM endpoints\n - server_metrics: Server metrics from Prometheus endpoints\n - runtime: Worker and communication settings\n - logging: Logging and debug settings\n - metrics: Metrics aggregation configuration\n - accuracy: Accuracy evaluation configuration\n\nValidation:\n The configuration is validated in several stages:\n 1. Individual field validation (types, ranges, formats)\n 2. Cross-field validation (mutual exclusivity, dependencies)\n\nEnvironment Variables:\n Values can reference environment variables using ${VAR} syntax.\n Optional defaults: ${VAR:default_value}\n\n Example:\n api_key: ${OPENAI_API_KEY}\n api_key: ${OPENAI_API_KEY:sk-default}","properties":{"models":{"description":"Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced. Accepts a model name string, list of model names, or full ModelsAdvanced config.","oneOf":[{"type":"string","description":"Single model name (normalized to ModelsAdvanced)."},{"type":"array","items":{"type":"string"},"minItems":1,"description":"List of model names (normalized to ModelsAdvanced)."},{"$ref":"#/$defs/ModelsAdvanced"}]},"endpoint":{"$ref":"#/$defs/EndpointConfig","description":"Endpoint configuration for connecting to inference server(s). Includes URLs, API type, authentication, timeout, and connection settings."},"datasets":{"description":"Dataset configuration as a single-element list. The list shape exists to share the schema between YAML and the AIPerfSweep CRD; the runtime currently loads exactly one dataset. Singular `dataset:` shorthand at the BenchmarkConfig top level is normalized to a one-entry list with name='default'.","items":{"discriminator":{"propertyName":"type","mapping":{"synthetic":"#/$defs/SyntheticDataset","file":"#/$defs/FileDataset","public":"#/$defs/PublicDataset"}},"anyOf":[{"$ref":"#/$defs/SyntheticDataset"},{"$ref":"#/$defs/FileDataset"},{"$ref":"#/$defs/PublicDataset"}]},"maxItems":1,"minItems":1,"title":"Datasets","type":"array"},"phases":{"description":"Ordered benchmark phases. Each entry must have a unique 'name' (e.g. 'warmup', 'profiling'). Order in the list IS the execution order; the first phase runs first. Single-config shorthand ({'type': 'concurrency', ...}) is normalized to a list of one. Top-level 'warmup:'/'profiling:' shorthand is normalized to a [warmup, profiling] list pre-validation. Shorthand: use a single phase config directly (with 'type' key) and it is normalized to a one-entry phases list.","oneOf":[{"description":"Ordered benchmark phases. Each entry must have a unique 'name' (e.g. 'warmup', 'profiling'). Order in the list IS the execution order; the first phase runs first. Single-config shorthand ({'type': 'concurrency', ...}) is normalized to a list of one. Top-level 'warmup:'/'profiling:' shorthand is normalized to a [warmup, profiling] list pre-validation.","items":{"discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}},"anyOf":[{"$ref":"#/$defs/ConcurrencyPhase"},{"$ref":"#/$defs/PoissonPhase"},{"$ref":"#/$defs/GammaPhase"},{"$ref":"#/$defs/ConstantPhase"},{"$ref":"#/$defs/UserCentricPhase"},{"$ref":"#/$defs/FixedSchedulePhase"}]},"minItems":1,"title":"Phases","type":"array"},{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Single phase config (normalized to a one-entry phases list).","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}}]},"artifacts":{"$ref":"#/$defs/ArtifactsConfig","description":"Artifacts configuration for benchmark output. Controls output directory and export formats."},"slos":{"anyOf":[{"additionalProperties":{"oneOf":[{"type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"type":"object"},{"type":"null"}],"default":null,"description":"SLO (Service Level Objectives) configuration as a generic dict. Maps metric names to threshold values. A request is counted as good only if it meets ALL specified thresholds.","title":"Slos","x-jinja2-supported":true},"tokenizer":{"anyOf":[{"$ref":"#/$defs/TokenizerConfig"},{"type":"null"}],"default":null,"description":"Tokenizer configuration for token counting. Used for ISL/OSL enforcement and accurate metrics. If not specified, uses the first model name."},"gpuTelemetry":{"$ref":"#/$defs/GpuTelemetryConfig","description":"GPU telemetry configuration for DCGM metrics collection. Collects GPU metrics (power, utilization, temperature) from DCGM endpoints. Enabled by default. Set enabled: false to disable."},"serverMetrics":{"$ref":"#/$defs/ServerMetricsConfig","description":"Server metrics configuration for Prometheus scraping. Collects operational metrics (queue depth, KV cache, batch sizes) from inference server Prometheus endpoints. Enabled by default. Set enabled: false to disable."},"networkLatency":{"$ref":"#/$defs/NetworkLatencyConfig","description":"Network latency calibration configuration. Probes the endpoint RTT during profiling and subtracts the mean from latency metrics, emitted as separate network_adjusted_* metrics. Disabled by default."},"otel":{"$ref":"#/$defs/OTelConfig"},"mlflow":{"$ref":"#/$defs/MLflowConfig"},"wandb":{"$ref":"#/$defs/WandbConfig"},"runtime":{"$ref":"#/$defs/RuntimeConfig","description":"Runtime configuration for worker processes and inter-process communication."},"logging":{"$ref":"#/$defs/LoggingConfig"},"metrics":{"$ref":"#/$defs/MetricsConfig","description":"Metrics aggregation configuration for benchmark summaries."},"accuracy":{"anyOf":[{"$ref":"#/$defs/AccuracyConfig"},{"type":"null"}],"default":null,"description":"Accuracy benchmarking configuration. When set, enables accuracy evaluation alongside performance profiling."},"scenario":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Lock all benchmark invariants for a named scenario (e.g. 'inferencex-agentx-mvp'). Plain data here; the lock is applied by the ScenarioResolver step in the pre-bootstrap resolver chain (auto-fills defaults, validates, raises ScenarioLockError on conflict). Distinct from the sweep ``scenarios`` strategy (SweepConfig), which expands hand-picked named runs -- this field is a single invariant LOCK, not a sweep.","title":"Scenario"},"unsafeOverride":{"default":false,"description":"Convert scenario lock errors to warnings; stamps submission_valid=false in the resolved scenario outcome. No-op without ``scenario``. Plain data; consumed by the ScenarioResolver.","title":"Unsafeoverride","type":"boolean"},"model":{"description":"Shorthand for 'models'. Accepts a single value which will be converted to a list/dict. Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced.","oneOf":[{"type":"string"},{"type":"array","items":{"type":"string"}},{"$ref":"#/$defs/ModelsAdvanced","description":"Model configuration. Accepts a single model name string, a list of model names, or an advanced configuration with strategy and weighted items. All forms are normalized to ModelsAdvanced.","x-kubernetes-preserve-unknown-fields":true}],"x-singular-alias-of":"models"},"dataset":{"anyOf":[{"additionalProperties":false,"description":"Synthetic dataset configuration.\n\nGenerates prompts programmatically based on token length\nspecifications. Ideal for controlled experiments.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"synthetic","description":"Dataset type discriminator. Must be 'synthetic'.","title":"Type","type":"string"},"entries":{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","minimum":1,"title":"Entries","type":"integer"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic dataset generation. When set, makes synthetic prompts, sampling, and other random operations reproducible across runs. Essential for A/B testing and debugging. Overrides global random_seed for this dataset.","title":"Randomseed"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptConfig"},{"type":"null"}],"default":null,"description":"Prompt/token length configuration specifying ISL, OSL, sequence distributions, and batch processing settings."},"prefixPrompts":{"anyOf":[{"$ref":"#/$defs/PrefixPromptConfig"},{"type":"null"}],"default":null,"description":"Shared prefix configuration for KV cache testing. Generates prefix prompts that are prepended to user prompts, simulating cached context scenarios."},"turns":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Number of request-response turns per conversation. Can be a fixed integer or {mean, stddev} distribution. Each turn consists of a user message and model response. Set to 1 for single-turn interactions. Multi-turn conversations enable testing of context retention and conversation history handling.","title":"Turns"},"turnDelay":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Delay in milliseconds between consecutive turns within a multi-turn conversation. Can be a fixed value or {mean, stddev} distribution. Simulates user think time between receiving a response and sending the next message. Only applies when turns > 1. Set to 0 for back-to-back turns.","title":"Turndelay"},"turnDelayRatio":{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","minimum":0.0,"title":"Turndelayratio","type":"number"},"images":{"anyOf":[{"$ref":"#/$defs/ImageConfig"},{"type":"null"}],"default":null,"description":"Synthetic image configuration for multimodal vision-language testing."},"audio":{"anyOf":[{"$ref":"#/$defs/AudioConfig"},{"type":"null"}],"default":null,"description":"Synthetic audio configuration for multimodal speech/audio testing."},"video":{"anyOf":[{"$ref":"#/$defs/VideoConfig"},{"type":"null"}],"default":null,"description":"Synthetic video configuration for multimodal video understanding testing."},"rankings":{"anyOf":[{"$ref":"#/$defs/RankingsConfig"},{"type":"null"}],"default":null,"description":"Rankings/reranking configuration for generating query-passage pairs. Only relevant for rankings endpoint types."},"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.isl; hoisted into prompts before validation.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.osl; hoisted into prompts before validation.","title":"Osl"}},"required":["type"],"title":"SyntheticDataset","type":"object"},{"additionalProperties":false,"description":"File-based dataset configuration.\n\nLoads prompts from a local file in various formats.\nSupports trace replay and custom sampling strategies.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"file","description":"Dataset type discriminator. Must be 'file'.","title":"Type","type":"string"},"path":{"anyOf":[{"format":"path","type":"string"},{"$ref":"#/$defs/RemoteRef"},{"type":"null"}],"default":null,"description":"Path to file or directory containing benchmark dataset, or a remote reference: `hf://[datasets/]/[@][/]`, `s3:///[?versionId=]`, or an `https://`/`http://` URL. Can be absolute or relative. Mutually exclusive with `records:`. Supported formats depend on the format field: JSONL for single_turn/multi_turn, JSONL (optionally gzipped) for tracelab, JSONL trace files for mooncake_trace/bailian_trace, Parquet for baseten_trace, directories for random_pool.","title":"Path"},"records":{"anyOf":[{"items":{"additionalProperties":true,"type":"object"},"type":"array"},{"additionalProperties":{"items":{"additionalProperties":true,"type":"object"},"type":"array"},"type":"object"},{"type":"null"}],"default":null,"description":"Inline benchmark records, embedded directly in the YAML config. Mutually exclusive with `path:`. The element schema is determined by `format:` (same shape as one line of the equivalent JSONL file). For `format: random_pool`, may be either a flat list (single pool) or a dict-of-lists (multi-pool, mirrors the directory-of-JSONLs file mode). All other formats require a flat list.","title":"Records"},"format":{"$ref":"#/$defs/DatasetFormat","default":"single_turn","description":"Dataset file format determining parsing logic and expected file structure. single_turn: JSONL with single prompt-response exchanges. multi_turn: JSONL with conversation history. tracelab: TraceLab agentic-coding corpus, JSONL or gzipped JSONL. mooncake_trace / bailian_trace / baseten_trace / burst_gpt_trace: timestamped trace files for replay. sagemaker_data_capture: JSONL captured by SageMaker DataCapture. random_pool: directory of reusable prompts."},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration. Allows scaling timestamps and token lengths before replay. Applies to trace formats such as mooncake_trace and baseten_trace, except speedup_ratio, which is rejected for baseten_trace (use replay_speedup / --replay-speedup to scale replay pacing there)."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Limit number of records to use from file. If not specified, uses all records in the file.","title":"Entries"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic sampling. When set, makes random/shuffle sampling reproducible across runs. Overrides global random_seed for this dataset.","title":"Randomseed"},"traceSessionSampleRatio":{"anyOf":[{"exclusiveMinimum":0.0,"maximum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Fraction of trace sessions to keep for replay, sampled whole-session to preserve multi-turn integrity; deterministic when ``random_seed`` is set. Only supported by the baseten_trace loader.","title":"Tracesessionsampleratio"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds; ``None`` disables the cap. Honored by the DAG JSONL loader and the baseten_trace loader's closed-loop think-times; the clamp count is reported at end of load.","title":"Interturndelaycapseconds"},"maxIdleGapCapSeconds":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Collapse idle gaps between consecutive requests (across all sessions) to at most this many seconds, so a sparse or session-sampled trace does not replay dead air; ``None`` disables the cap. The cap is in replay wall-clock seconds, applied after replay_speedup compression, so it bounds actual benchmark idle time regardless of speedup. Only supported by the baseten_trace loader.","title":"Maxidlegapcapseconds"},"replaySpeedup":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Trace replay wall-clock compression (10 = 10x faster than recorded): divides normalized timestamps and inter-turn delays; ``None`` = real time. Unlike synthesis speedup_ratio, hash_ids stay untouched (KV-cache fidelity). Only supported by the baseten_trace loader.","title":"Replayspeedup"},"openLoopReplay":{"default":true,"description":"Open-loop replay (the default): each session starts at its absolute, speedup-scaled recorded timestamp; continuation turns fire at max(recorded timestamp, prior-turn completion). Set ``False`` for closed-loop back-pressure: continuation turns fire a think-time (recorded start-to-start gap minus recorded e2e duration) after the prior turn completes, keeping sessions causally ordered when replayed service times differ from recorded (e.g. A/A comparisons). Only honored by the baseten_trace loader.","title":"Openloopreplay","type":"boolean"},"openLoopStrict":{"default":false,"description":"In open-loop replay, fire every trace row at its absolute recorded timestamp as an independent single-turn session, trading away multi-turn grouping and session metrics. Only honored by the baseten_trace loader.","title":"Openloopstrict","type":"boolean"},"omitKvHints":{"default":false,"description":"Drop recorded KV-cache hints (``hash_ids``, ``block_size``) from replayed request bodies, for strict frontends that reject unknown parameters. Only honored by the baseten_trace loader.","title":"Omitkvhints","type":"boolean"},"forceMinTokens":{"default":true,"description":"Pin ``min_tokens`` to the recorded output length so replayed generations match recorded lengths; disable to let EOS end generations naturally (some servers reject ``min_tokens``). Only honored by the baseten_trace loader.","title":"Forcemintokens","type":"boolean"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one. Can be a fixed integer or {mean, stddev} distribution. Per-line `output_length` values in the file always take precedence.","title":"Osl"},"promptBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of text items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable text inputs entirely (e.g. image/audio/video-only workloads).","title":"Promptbatchsize"},"imageBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of images per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable image inputs entirely.","title":"Imagebatchsize"},"audioBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of audio items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable audio inputs entirely.","title":"Audiobatchsize"},"videoBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Number of video items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable video inputs entirely.","title":"Videobatchsize"},"blockSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Hash-id block granularity for trace replay (--isl-block-size). hash-id trace loaders (mooncake_trace, bailian_trace, baseten_trace, tracelab) decode each hash_id into a cached block of this many tokens; total ISL = (num_hash_ids - 1) * block_size + final_block_size. When unset, the loader's plugin-metadata default applies (e.g. 512 for mooncake_trace, 16 for bailian_trace). Not used by weka, which carries its own inline per-block sizes.","title":"Blocksize"},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree. The idle clock covers initial future work and restarts when the final in-flight request across the root and all descendant streams completes. If no request from that tree reaches the wire before the cap, AIPerf uniformly advances only that tree's pending replay timers. Dataset timestamps are unchanged, and request order, spawn/join dependencies, and replay barriers remain authoritative. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds"},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from trace datasets at load time. With this flag, Turn.timestamp and Turn.delay are emitted as None so concurrency / request-rate timing modes dispatch turns back-to-back instead of reproducing the recorded user think-time gaps. No effect under fixed-schedule (timestamps drive that mode before they could be ignored -- combine with --no-fixed-schedule if you want both behaviors). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For weka_trace inputs, emit Turn.delay using only the recorded per-request `think_time` (client-side delay before each request) instead of the full `t_curr - t_prev` inter-request delta. Compresses replay wall time against zero-latency mocks because the recorded `api_time` portion of each gap is dropped. Mirrors kv-cache-tester's default `--timing-strategy think-only`. Falls back to the full delta for turns whose recorded `think_time` is null. Mutually exclusive with ignore_trace_delays. No effect on non-weka trace loaders.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per Weka root trace. Only honored by format weka_trace: the loader drops whole traces whose *recorded* peak exceeds this ceiling at load time (filter-then-cap before entries). Not a DatasetManager tokenize filter; rejected for non-Weka formats.","title":"Maxcontextlength"},"remoteFixedScheduleCheck":{"anyOf":[{"$ref":"#/$defs/RemoteFixedScheduleCheck"},{"type":"null"}],"default":null,"description":"Internal: set by the CLI for a remote `--input-file` and cleared by the dataset preflight once the first record decides fixed-schedule auto-promotion. Do not set in YAML."},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for trace replay (mirror of the synthetic dataset's prompts.cache_bust). The engine reads the active dataset's target via ``BenchmarkConfig.get_cache_bust_target()``; 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."}},"required":["type"],"title":"FileDataset","type":"object"},{"additionalProperties":false,"description":"Public dataset configuration.\n\nUses well-known public benchmarking datasets that are\nautomatically downloaded and processed by AIPerf.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"public","description":"Dataset type discriminator. Must be 'public'.","title":"Type","type":"string"},"dataset":{"$ref":"#/$defs/PublicDatasetType","description":"Pre-configured public dataset to download and use for benchmarking. Name of the HuggingFace public dataset enum (e.g. 'sharegpt', 'alpaca'). AIPerf automatically downloads and parses these datasets."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Limit number of records to use from the dataset. If not specified, uses all available records.","title":"Entries"},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"}],"default":null,"description":"Random seed for deterministic sampling from the dataset. Overrides global random_seed for this dataset.","title":"Randomseed"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"hfSubset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset subset/config name override (e.g. 'sharegpt4o'). Only applies for HuggingFace-backed public dataset loaders. Takes priority over the subset defined in the plugin registry.","title":"Hfsubset"},"filters":{"additionalProperties":{"type":"string"},"description":"Dataset-specific filters forwarded to public dataset loaders. Supported keys and values depend on the selected dataset.","title":"Filters","type":"object"},"hfMooncakeDataset":{"anyOf":[{"minLength":1,"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace repository for mooncake_hf; required for this loader. The CLI flag --hf-mooncake-dataset auto-selects this loader.","title":"Hfmooncakedataset"},"hfMooncakeSplit":{"anyOf":[{"minLength":1,"pattern":"\\S","type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace split override for mooncake_hf (e.g. train or test). Unset uses the plugin default, test. Only supported by mooncake_hf.","title":"Hfmooncakesplit"},"mooncakeTraceToHfMapping":{"additionalProperties":{"type":"string"},"description":"Mooncake target field to dotted source column path for mooncake_hf. Supports messages, session_id, tools, output_length, extra, and extra.PARAMETER.","title":"Mooncaketracetohfmapping","type":"object"},"hfWekaDataset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset repo for the generic Weka loader (e.g. `semianalysisai/cc-traces-weka-061526`). Only applies with `dataset: weka_hf`; setting it with any other public dataset is an error. Passing `--hf-weka-dataset` on the CLI auto-selects `--public-dataset weka_hf`, so the repo flag works on its own. Pinned Weka public dataset aliases keep their registry-defined repo names.","title":"Hfwekadataset"},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds for HF-backed Weka trace replay (mirror of ``FileDataset.inter_turn_delay_cap_seconds``). ``None`` disables the cap.","title":"Interturndelaycapseconds"},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree for HF-backed Weka replay (mirror of ``FileDataset.trace_idle_gap_cap_seconds``). Dataset timestamps remain unchanged. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds"},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from HF-backed Weka trace replay at load time (mirror of ``FileDataset.ignore_trace_delays``). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For HF-backed Weka replay, emit Turn.delay using only the recorded per-request ``think_time`` instead of the full inter-request delta (mirror of ``FileDataset.use_think_time_only``). Mutually exclusive with ignore_trace_delays.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per conversation for HF-backed Weka replay; drops over-length traces at load using *recorded* lengths (mirror of ``FileDataset.max_context_length``). Rejected for non-Weka public datasets.","title":"Maxcontextlength"},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for HF-backed Weka trace replay (mirror of ``FileDataset.cache_bust``). 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration for HF-backed Weka trace replay (mirror of ``FileDataset.synthesis``). ``--max-isl``/``--max-osl`` cap the per-conversation input/output lengths of the replayed HF Weka traces; without this they applied only to file-based traces. ``None`` disables synthesis."},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one (mirror of ``FileDataset.osl``). Can be a fixed integer or {mean, stddev} distribution. Per-record output lengths always take precedence.","title":"Osl"}},"required":["type","dataset"],"title":"PublicDataset","type":"object"}],"description":"Single dataset config (normalized to a one-entry datasets list).","x-singular-alias-of":"datasets"},"warmup":{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Warmup phase shorthand. Requires profiling when used.","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}},"profiling":{"anyOf":[{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConcurrencyPhase","type":"object"},{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate"},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}."},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration"},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod"},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax"},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests"},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration"},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions"},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency"},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency"},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}."},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod"},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold"},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds"},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration"},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane"},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod"},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset"},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset"},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["type"],"title":"FixedSchedulePhase","type":"object"}],"description":"Profiling phase shorthand. Normalized to a phases entry named 'profiling'.","discriminator":{"propertyName":"type","mapping":{"concurrency":"#/$defs/ConcurrencyPhase","poisson":"#/$defs/PoissonPhase","gamma":"#/$defs/GammaPhase","constant":"#/$defs/ConstantPhase","user_centric":"#/$defs/UserCentricPhase","fixed_schedule":"#/$defs/FixedSchedulePhase"}}}},"required":["endpoint"],"title":"BenchmarkConfig","type":"object","allOf":[{"anyOf":[{"required":["model"]},{"required":["models"]}]},{"anyOf":[{"required":["dataset"]},{"required":["datasets"]}]},{"anyOf":[{"required":["phases"]},{"required":["profiling"]}]}]},"CacheBustConfig":{"additionalProperties":false,"description":"Per-run cache-bust marker configuration.\n\nTwo families of targets:\n\nRID targets (``system_prefix``, ``system_suffix``, ``first_turn_prefix``,\n``first_turn_suffix``): inject a per-trajectory unique marker derived from a\nSHA-256 digest of (benchmark_id, recycle_pass, trajectory_index, trace_id).\nThe digest is phase-agnostic — warmup and profiling turns for the same\ntrajectory share the same marker — so warmup KV-cache work transfers to\nprofiling. Different trajectories (different lanes or recycle passes) receive\ndistinct markers. Prefix variants diverge at token 0; suffix variants append\nafter existing content.\n\nWarmup-isolation targets (``warmup_isolation_system``,\n``warmup_isolation_first_turn``): inject a constant ``[warmup]`` marker only\nduring the WARMUP phase; during PROFILING the marker is ``None`` so the\nprofiling request arrives clean. These targets are phase-aware and do NOT use\na per-trajectory digest. Incompatible with ``agentic_replay`` timing mode.\n\n``media_uuids`` extends a RID target to media: the ``uuid`` of every media\ncontent part gets the trajectory's digest as a suffix, so a replayed\ntrajectory cannot reuse the server's multimodal cache entries for the same\nmedia either.","properties":{"target":{"$ref":"#/$defs/CacheBustTarget","default":"none","description":"Where (and how) to inject a per-conversation cache-bust marker. Prefix variants prepend at token 0 (most aggressive); suffix variants append after existing content. 'none' disables the feature (default)."},"mediaUuids":{"default":false,"description":"Append the trajectory's RID digest to every media `uuid`, so a replay cannot hit vLLM's multimodal cache entries from an earlier play. Requires a RID target and the chat endpoint.","title":"Mediauuids","type":"boolean"}},"title":"CacheBustConfig","type":"object"},"CacheBustTarget":{"description":"Where (and how) to inject a per-conversation cache-bust marker.\n\nPrefix variants diverge at token 0 of the prompt (most aggressive -- defeats\nKV-cache prefix matching for the entire prompt). Suffix variants append\nafter existing content (preserves leading-prefix caching).","enum":["none","system_prefix","system_suffix","first_turn_prefix","first_turn_suffix","warmup_isolation_system","warmup_isolation_first_turn"],"title":"CacheBustTarget","type":"string"},"CancellationConfig":{"additionalProperties":false,"description":"Configuration for request cancellation testing.\n\nEnables testing server behavior when clients cancel requests mid-flight.","properties":{"rate":{"description":"Percentage of requests to cancel (0-100). 0.5 means half a percent; 10.0 means 10%% of requests will be cancelled.","title":"Rate","oneOf":[{"description":"Percentage of requests to cancel (0-100). 0.5 means half a percent; 10.0 means 10%% of requests will be cancelled.","maximum":100.0,"minimum":0.0,"title":"Rate","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"delay":{"default":0.0,"description":"Seconds to wait after sending before cancelling. 0.0 means cancel immediately after send.","title":"Delay","oneOf":[{"default":0.0,"description":"Seconds to wait after sending before cancelling. 0.0 means cancel immediately after send.","minimum":0.0,"title":"Delay","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["rate"],"title":"CancellationConfig","type":"object"},"ConcurrencyPhase":{"additionalProperties":false,"description":"Concurrency-controlled load: dispatch immediately when a slot opens.\n\nPrimary control is ``concurrency`` (defaults to 1).\nNo rate limiting -- pure concurrency-based throughput.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"concurrency","description":"Concurrency-controlled immediate dispatch.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","title":"Concurrency","oneOf":[{"default":1,"description":"Max concurrent in-flight requests (must be >= 1). Primary control for concurrency phases.","minimum":1,"title":"Concurrency","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"ConcurrencyPhase","type":"object"},"ConnectionReuseStrategy":{"description":"Transport connection reuse strategy. Controls how and when connections are reused across requests.","enum":["pooled","never","sticky-user-sessions"],"title":"ConnectionReuseStrategy","type":"string"},"ConstantPhase":{"additionalProperties":false,"description":"Constant-rate request arrivals (fixed inter-arrival time).","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"constant","description":"Constant rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"ConstantPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"ConvergenceConfig":{"additionalProperties":false,"description":"Adaptive trial-stopping criterion.\n\nPresence of this object on `MultiRunConfig.convergence` enables\nadaptive stopping: trials run until the criterion fires (or\n`MultiRunConfig.num_runs` is reached, whichever comes first).","properties":{"metric":{"description":"Metric tag to evaluate convergence on (e.g., 'ttft').","title":"Metric","type":"string"},"stat":{"$ref":"#/$defs/ConvergenceStat","default":"avg","description":"Statistic on the metric to inspect (avg/p50/p90/p95/p99)."},"mode":{"$ref":"#/$defs/ConvergenceCriterionType","default":"ci_width","description":"Convergence criterion plugin name (ci_width|cv|distribution)."},"threshold":{"anyOf":[{"exclusiveMaximum":1,"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Criterion threshold (interpretation depends on `mode`). When None (default), each criterion uses its own algorithm-specific default: ci_width=0.10, cv=0.05, distribution=0.05 (KS p-value).","title":"Threshold","x-jinja2-supported":true},"minRuns":{"default":2,"description":"Minimum trials before the criterion is checked.","title":"Minruns","oneOf":[{"default":2,"description":"Minimum trials before the criterion is checked.","minimum":2,"title":"Minruns","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metric"],"title":"ConvergenceConfig","type":"object"},"ConvergenceCriterionType":{"enum":["ci_width","cv","distribution"],"title":"ConvergenceCriterionType","type":"string"},"ConvergenceStat":{"description":"Statistic to evaluate for convergence when using ci_width or cv mode.","enum":["avg","p50","p90","p95","p99","min","max"],"title":"ConvergenceStat","type":"string"},"DatasetFormat":{"enum":["burst_gpt_trace","bailian_trace","baseten_trace","mooncake_trace","h_cua_perf","raw_payload","inputs_json","dag_jsonl","sagemaker_data_capture","multi_turn","random_pool","single_turn","speed_bench_qualitative","speed_bench_throughput_1k","speed_bench_throughput_2k","speed_bench_throughput_8k","speed_bench_throughput_16k","speed_bench_throughput_32k","speed_bench_throughput_100k","tracelab","weka_trace"],"title":"DatasetFormat","type":"string"},"DatasetSamplingStrategy":{"enum":["random","sequential","shuffle"],"title":"DatasetSamplingStrategy","type":"string"},"DualBindCommunicationConfig":{"additionalProperties":false,"description":"Dual-bind (IPC + TCP) communication configuration.\n\nFor Kubernetes deployments where controller services connect via IPC\n(co-located in same pod) and worker pods connect via TCP.\n\nWhen controller_host is None, services use IPC (local mode).\nWhen controller_host is set, services use TCP to that host (remote mode).","properties":{"type":{"const":"dual","description":"Communication type. Must be 'dual'.","title":"Type","type":"string"},"ipcPath":{"default":"/tmp/aiperf","description":"Directory for IPC socket files (local services).","title":"Ipcpath","type":"string"},"tcpHost":{"default":"0.0.0.0","description":"TCP bind host for proxies. Use 0.0.0.0 to listen on all interfaces.","title":"Tcphost","type":"string"},"controllerHost":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Controller host for remote TCP connections. When set, services connect via TCP to this host instead of IPC. In Kubernetes, set via JobSet DNS (e.g., controller.namespace.svc).","title":"Controllerhost"},"recordsPort":{"default":5557,"description":"Port for records push/pull communication.","title":"Recordsport","oneOf":[{"default":5557,"description":"Port for records push/pull communication.","maximum":65535,"minimum":1,"title":"Recordsport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditRouterPort":{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","title":"Creditrouterport","oneOf":[{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Creditrouterport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"controlPort":{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","title":"Controlport","oneOf":[{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Controlport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditReturnPushPullPort":{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","title":"Creditreturnpushpullport","oneOf":[{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","maximum":65535,"minimum":1,"title":"Creditreturnpushpullport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"eventBusProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Event bus proxy ports (XPUB/XSUB)."},"datasetManagerProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Dataset manager proxy ports (DEALER/ROUTER)."},"rawInferenceProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Raw inference proxy ports (PUSH/PULL)."}},"required":["type"],"title":"DualBindCommunicationConfig","type":"object"},"EmpiricalDistribution":{"additionalProperties":false,"description":"Discrete distribution sampled from weighted values.\n\nYAML:\n isl:\n points:\n - {value: 128, weight: 40}\n - {value: 512, weight: 35}\n - {value: 2048, weight: 20}\n - {value: 8192, weight: 5}","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"points":{"description":"Weighted discrete values. Weights are relative.","items":{"$ref":"#/$defs/EmpiricalPoint"},"title":"Points","type":"array"},"type":{"const":"empirical","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["points"],"title":"EmpiricalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"EmpiricalPoint":{"additionalProperties":false,"description":"A weighted value in an empirical distribution.","properties":{"value":{"description":"The discrete value.","title":"Value","oneOf":[{"description":"The discrete value.","title":"Value","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"weight":{"default":1.0,"description":"Relative weight (normalized internally).","title":"Weight","oneOf":[{"default":1.0,"description":"Relative weight (normalized internally).","exclusiveMinimum":0.0,"title":"Weight","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["value"],"title":"EmpiricalPoint","type":"object"},"EndpointConfig":{"additionalProperties":false,"description":"Endpoint configuration for connecting to inference servers.\n\nThis section configures how AIPerf connects to and communicates\nwith the target inference server(s). It supports multiple URLs\nfor load-balanced deployments and various API types.","properties":{"urls":{"description":"List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","items":{"type":"string"},"title":"Urls","type":"array","x-kubernetes-preserve-unknown-fields":true},"urlStrategy":{"$ref":"#/$defs/URLSelectionStrategy","default":"round_robin","description":"Strategy for distributing requests across multiple URLs. round_robin cycles through URLs in order."},"type":{"$ref":"#/$defs/EndpointType","default":"chat","description":"API endpoint type determining request/response format. chat: OpenAI chat completions, completions: OpenAI completions, embeddings: vector embeddings, rankings: reranking, template: custom format, and others — see `aiperf plugins` for the full list."},"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Override default endpoint path. Use for servers with non-standard API paths. Example: '/custom/v2/generate'","title":"Path"},"apiKey":{"anyOf":[{"type":"string"},{"type":"null"}],"description":"API authentication key. Supports environment variable substitution: ${OPENAI_API_KEY}. Can also use ${VAR:default} syntax for defaults.","title":"Apikey"},"timeout":{"default":21600,"description":"Request timeout in seconds (0 = no timeout). Requests exceeding this duration are marked as failed. Should exceed expected max response time.","title":"Timeout","oneOf":[{"default":21600,"description":"Request timeout in seconds (0 = no timeout). Requests exceeding this duration are marked as failed. Should exceed expected max response time.","minimum":0.0,"title":"Timeout","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"streaming":{"default":false,"description":"Enable streaming (Server-Sent Events) responses. Required for accurate TTFT (time to first token) measurement. Server must support streaming for this to work.","title":"Streaming","type":"boolean"},"allowEmptyContent":{"default":false,"description":"Retain explicitly empty content or reasoning strings as timed responses for streaming OpenAI Chat Completions. This changes the shared response timeline without adding output tokens.","title":"Allowemptycontent","type":"boolean"},"transport":{"anyOf":[{"$ref":"#/$defs/TransportType"},{"type":"null"}],"default":null,"description":"Transport plugin name. 'http' (aiohttp-based HTTP/1.1) and 'sagemaker' (AWS SageMaker Runtime, normally derived from sagemaker.endpoint_name) are shipped. Auto-detected from the URL when unset, which always selects 'http'; explicit setting overrides auto-detection."},"awsRegion":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"AWS region for the request. Required when auth_type='sigv4'.","title":"Awsregion"},"awsProfile":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Named AWS credentials profile. Unset uses botocore's default credential chain.","title":"Awsprofile"},"authType":{"anyOf":[{"$ref":"#/$defs/RequestSignerType"},{"type":"null"}],"default":null,"description":"Request signing method for authentication. When set, the selected request_signer plugin signs every HTTP request sent by the HTTP transport. Replaces Bearer token auth (api_key is ignored when auth_type is set)."},"awsService":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SigV4 signing name -- the credential scope the signature is bound to (e.g. 'execute-api', 'sagemaker', 'bedrock'). This is not always the API id: the 'sagemaker-runtime' API signs as 'sagemaker' and 'bedrock-runtime' as 'bedrock'. Required when auth_type='sigv4', unless the selected transport declares which AWS API it speaks, in which case the scope is resolved from botocore's service model.","title":"Awsservice"},"sagemaker":{"$ref":"#/$defs/SageMakerConfig","description":"SageMaker Runtime routing options. Setting sagemaker.endpoint_name selects the SageMaker transport."},"connectionReuse":{"$ref":"#/$defs/ConnectionReuseStrategy","default":"pooled","description":"HTTP connection management strategy. pooled: shared connection pool (fastest), never: new connection per request (includes TCP overhead), sticky-user-sessions: dedicated connection per user session."},"useLegacyMaxTokens":{"default":false,"description":"Use 'max_tokens' field instead of 'max_completion_tokens'. Enable for compatibility with older OpenAI API versions.","title":"Uselegacymaxtokens","type":"boolean"},"useServerTokenCount":{"default":false,"description":"Use server-reported token counts from response usage field. When true, trusts usage.prompt_tokens and usage.completion_tokens. When false, counts tokens locally using configured tokenizer.","title":"Useservertokencount","type":"boolean"},"perChunkUsage":{"default":false,"description":"Request per-chunk token usage on streaming responses by setting stream_options.continuous_usage_stats. When true, the server reports cumulative usage on every chunk (not just the final one), which lets inter-token latency subtract the first content chunk's real token count instead of assuming one token. Requires a server that supports continuous_usage_stats (e.g. vLLM, TRT-LLM); strict OpenAI rejects it. Requires use_server_token_count so OSL and the first-chunk count share the server source.","title":"Perchunkusage","type":"boolean"},"template":{"anyOf":[{"$ref":"#/$defs/TemplateConfig"},{"type":"null"}],"default":null,"description":"Custom template configuration for template endpoint type. Only used when type='template'. Defines request body format and response parsing."},"headers":{"additionalProperties":{"type":"string"},"description":"Custom HTTP headers to include in all requests. Useful for authentication, tracing, or routing. Values support environment variable substitution.","title":"Headers","type":"object"},"urlHeaders":{"description":"Extra HTTP headers for individual targets: entry ``i`` applies to requests sent to ``urls[i]`` and may be ``{}``. Merged over ``headers``; on a name clash (case-insensitive) the per-URL value wins. Use for routing hints such as ``X-data-parallel-rank``. Credential-bearing header names are rejected; put credentials in ``headers`` or ``api_key``. Example: [{X-data-parallel-rank: '0'}, {X-data-parallel-rank: '1'}]","items":{"additionalProperties":{"type":"string"},"type":"object"},"title":"Urlheaders","type":"array"},"extra":{"additionalProperties":true,"description":"Additional fields to include in request body. Merged into every request. Common fields: temperature, top_p, top_k, stop.","title":"Extra","type":"object"},"downloadVideoContent":{"default":false,"description":"For video generation endpoints, download the video content after generation completes. Adds a content download step to the async polling flow.","title":"Downloadvideocontent","type":"boolean"},"requestContentType":{"anyOf":[{"$ref":"#/$defs/RequestContentType"},{"type":"null"}],"default":null,"description":"Content type for request body serialization. Default is 'application/json'. Set to 'multipart/form-data' for servers that require form-encoded requests (e.g. vLLM video generation)."},"sessionHeader":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HTTP header name used to carry the per-session affinity identifier. When set, replaces the default `X-Correlation-ID` header. Useful when the inference server expects a custom session-affinity header (e.g. `--session-header X-Session-ID`).","title":"Sessionheader"},"uuidAndStrip":{"default":false,"description":"Enable AIPerf-managed image stripping for vLLM's multimodal processor cache. Dataset-authored image UUIDs, including UUID-only cache references, always pass through on the chat endpoint; this flag only strips repeated content after AIPerf observes it in the same session. Automatic stripping supports only single_turn datasets with session_id-grouped rows; multi_turn is rejected. The server cache must cover the working set, and requests in a session must reach a replica that retains earlier UUIDs.","title":"Uuidandstrip","type":"boolean"},"waitForModelTimeout":{"default":0.0,"description":"Enable a pre-flight readiness probe by setting this to a positive value (seconds). aiperf applies this timeout to each URL/model probe before starting the benchmark, aborting with a non-zero exit if any probe times out. For multiple URLs or models, worst-case wall-clock time can be roughly this timeout multiplied by the number of URL/model probes. The probe strategy is controlled by `--wait-for-model-mode`, which defaults to sending a 1-token inference request. 0 (default) disables the probe. Eliminates the need for external shell-based readiness loops in containers and Kubernetes recipes.","title":"Waitformodeltimeout","oneOf":[{"default":0.0,"description":"Enable a pre-flight readiness probe by setting this to a positive value (seconds). aiperf applies this timeout to each URL/model probe before starting the benchmark, aborting with a non-zero exit if any probe times out. For multiple URLs or models, worst-case wall-clock time can be roughly this timeout multiplied by the number of URL/model probes. The probe strategy is controlled by `--wait-for-model-mode`, which defaults to sending a 1-token inference request. 0 (default) disables the probe. Eliminates the need for external shell-based readiness loops in containers and Kubernetes recipes.","minimum":0.0,"title":"Waitformodeltimeout","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"waitForModelInterval":{"default":5.0,"description":"Seconds between readiness probe attempts. Only consulted when `--wait-for-model-timeout` is positive.","title":"Waitformodelinterval","oneOf":[{"default":5.0,"description":"Seconds between readiness probe attempts. Only consulted when `--wait-for-model-timeout` is positive.","exclusiveMinimum":0.0,"title":"Waitformodelinterval","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"waitForModelMode":{"default":"inference","description":"Strategy for the readiness probe. 'inference' (default): POST a 1-token inference request to the configured endpoint; this is the strongest signal — it proves the full stack (frontend, scheduler, worker, forward pass) is live. Any HTTP status < 500 counts as ready. 'models': GET `/v1/models` and verify the model id appears in `data[]` (cheaper, no tokens consumed; falls back to a plain GET on the base URL on 404). 'both': run 'models' first, then 'inference'. Only consulted when `--wait-for-model-timeout` is positive.","enum":["models","inference","both"],"title":"Waitformodelmode","type":"string"},"resetKvCache":{"anyOf":[{"type":"boolean"},{"$ref":"#/$defs/ResetKvCacheConfig"},{"type":"null"}],"default":null,"description":"When enabled, POST a KV-cache reset once per logical benchmark cell before warmup/profiling. Accepts false | true | {timeout_seconds?, path?}."},"serverProfiler":{"anyOf":[{"type":"boolean"},{"$ref":"#/$defs/ServerProfilerConfig"},{"type":"null"}],"default":null,"description":"When enabled, start/stop the server profiler around each profiling phase. Accepts false | true | {timeout_seconds?, start_path?, stop_path?}."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","oneOf":[{"type":"string"},{"description":"List of server URLs to benchmark. Requests distributed according to url_strategy. URLs without a scheme have ``http://`` prepended automatically. Example: ['http://localhost:8000/v1/chat/completions']","items":{"type":"string"},"title":"Urls","type":"array","x-kubernetes-preserve-unknown-fields":true}],"x-singular-alias-of":"urls"}},"required":[],"title":"EndpointConfig","type":"object","allOf":[{"anyOf":[{"required":["url"]},{"required":["urls"]}]}]},"EndpointType":{"enum":["audio_transcription","chat","cohere_rankings","completions","responses","messages","chat_embeddings","embeddings","hf_tei_rankings","huggingface_generate","image_generation","image_edit","video_generation","image_retrieval","nim_embeddings","nim_rankings","solido_rag","raw","template"],"title":"EndpointType","type":"string"},"FileDataset":{"additionalProperties":false,"description":"File-based dataset configuration.\n\nLoads prompts from a local file in various formats.\nSupports trace replay and custom sampling strategies.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"file","description":"Dataset type discriminator. Must be 'file'.","title":"Type","type":"string"},"path":{"anyOf":[{"format":"path","type":"string"},{"$ref":"#/$defs/RemoteRef"},{"type":"null"}],"default":null,"description":"Path to file or directory containing benchmark dataset, or a remote reference: `hf://[datasets/]/[@][/]`, `s3:///[?versionId=]`, or an `https://`/`http://` URL. Can be absolute or relative. Mutually exclusive with `records:`. Supported formats depend on the format field: JSONL for single_turn/multi_turn, JSONL (optionally gzipped) for tracelab, JSONL trace files for mooncake_trace/bailian_trace, Parquet for baseten_trace, directories for random_pool.","title":"Path"},"records":{"anyOf":[{"items":{"additionalProperties":true,"type":"object"},"type":"array"},{"additionalProperties":{"items":{"additionalProperties":true,"type":"object"},"type":"array"},"type":"object"},{"type":"null"}],"default":null,"description":"Inline benchmark records, embedded directly in the YAML config. Mutually exclusive with `path:`. The element schema is determined by `format:` (same shape as one line of the equivalent JSONL file). For `format: random_pool`, may be either a flat list (single pool) or a dict-of-lists (multi-pool, mirrors the directory-of-JSONLs file mode). All other formats require a flat list.","title":"Records"},"format":{"$ref":"#/$defs/DatasetFormat","default":"single_turn","description":"Dataset file format determining parsing logic and expected file structure. single_turn: JSONL with single prompt-response exchanges. multi_turn: JSONL with conversation history. tracelab: TraceLab agentic-coding corpus, JSONL or gzipped JSONL. mooncake_trace / bailian_trace / baseten_trace / burst_gpt_trace: timestamped trace files for replay. sagemaker_data_capture: JSONL captured by SageMaker DataCapture. random_pool: directory of reusable prompts."},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration. Allows scaling timestamps and token lengths before replay. Applies to trace formats such as mooncake_trace and baseten_trace, except speedup_ratio, which is rejected for baseten_trace (use replay_speedup / --replay-speedup to scale replay pacing there)."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Limit number of records to use from file. If not specified, uses all records in the file.","title":"Entries","x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic sampling. When set, makes random/shuffle sampling reproducible across runs. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"traceSessionSampleRatio":{"anyOf":[{"exclusiveMinimum":0.0,"maximum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Fraction of trace sessions to keep for replay, sampled whole-session to preserve multi-turn integrity; deterministic when ``random_seed`` is set. Only supported by the baseten_trace loader.","title":"Tracesessionsampleratio","x-jinja2-supported":true},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds; ``None`` disables the cap. Honored by the DAG JSONL loader and the baseten_trace loader's closed-loop think-times; the clamp count is reported at end of load.","title":"Interturndelaycapseconds","x-jinja2-supported":true},"maxIdleGapCapSeconds":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Collapse idle gaps between consecutive requests (across all sessions) to at most this many seconds, so a sparse or session-sampled trace does not replay dead air; ``None`` disables the cap. The cap is in replay wall-clock seconds, applied after replay_speedup compression, so it bounds actual benchmark idle time regardless of speedup. Only supported by the baseten_trace loader.","title":"Maxidlegapcapseconds","x-jinja2-supported":true},"replaySpeedup":{"anyOf":[{"exclusiveMinimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Trace replay wall-clock compression (10 = 10x faster than recorded): divides normalized timestamps and inter-turn delays; ``None`` = real time. Unlike synthesis speedup_ratio, hash_ids stay untouched (KV-cache fidelity). Only supported by the baseten_trace loader.","title":"Replayspeedup","x-jinja2-supported":true},"openLoopReplay":{"default":true,"description":"Open-loop replay (the default): each session starts at its absolute, speedup-scaled recorded timestamp; continuation turns fire at max(recorded timestamp, prior-turn completion). Set ``False`` for closed-loop back-pressure: continuation turns fire a think-time (recorded start-to-start gap minus recorded e2e duration) after the prior turn completes, keeping sessions causally ordered when replayed service times differ from recorded (e.g. A/A comparisons). Only honored by the baseten_trace loader.","title":"Openloopreplay","type":"boolean"},"openLoopStrict":{"default":false,"description":"In open-loop replay, fire every trace row at its absolute recorded timestamp as an independent single-turn session, trading away multi-turn grouping and session metrics. Only honored by the baseten_trace loader.","title":"Openloopstrict","type":"boolean"},"omitKvHints":{"default":false,"description":"Drop recorded KV-cache hints (``hash_ids``, ``block_size``) from replayed request bodies, for strict frontends that reject unknown parameters. Only honored by the baseten_trace loader.","title":"Omitkvhints","type":"boolean"},"forceMinTokens":{"default":true,"description":"Pin ``min_tokens`` to the recorded output length so replayed generations match recorded lengths; disable to let EOS end generations naturally (some servers reject ``min_tokens``). Only honored by the baseten_trace loader.","title":"Forcemintokens","type":"boolean"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one. Can be a fixed integer or {mean, stddev} distribution. Per-line `output_length` values in the file always take precedence.","title":"Osl"},"promptBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of text items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable text inputs entirely (e.g. image/audio/video-only workloads).","title":"Promptbatchsize","x-jinja2-supported":true},"imageBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of images per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable image inputs entirely.","title":"Imagebatchsize","x-jinja2-supported":true},"audioBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of audio items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable audio inputs entirely.","title":"Audiobatchsize","x-jinja2-supported":true},"videoBatchSize":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of video items per request. Only applies to format: random_pool; rejected on other formats. Set to 0 to disable video inputs entirely.","title":"Videobatchsize","x-jinja2-supported":true},"blockSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hash-id block granularity for trace replay (--isl-block-size). hash-id trace loaders (mooncake_trace, bailian_trace, baseten_trace, tracelab) decode each hash_id into a cached block of this many tokens; total ISL = (num_hash_ids - 1) * block_size + final_block_size. When unset, the loader's plugin-metadata default applies (e.g. 512 for mooncake_trace, 16 for bailian_trace). Not used by weka, which carries its own inline per-block sizes.","title":"Blocksize","x-jinja2-supported":true},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree. The idle clock covers initial future work and restarts when the final in-flight request across the root and all descendant streams completes. If no request from that tree reaches the wire before the cap, AIPerf uniformly advances only that tree's pending replay timers. Dataset timestamps are unchanged, and request order, spawn/join dependencies, and replay barriers remain authoritative. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds","x-jinja2-supported":true},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from trace datasets at load time. With this flag, Turn.timestamp and Turn.delay are emitted as None so concurrency / request-rate timing modes dispatch turns back-to-back instead of reproducing the recorded user think-time gaps. No effect under fixed-schedule (timestamps drive that mode before they could be ignored -- combine with --no-fixed-schedule if you want both behaviors). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For weka_trace inputs, emit Turn.delay using only the recorded per-request `think_time` (client-side delay before each request) instead of the full `t_curr - t_prev` inter-request delta. Compresses replay wall time against zero-latency mocks because the recorded `api_time` portion of each gap is dropped. Mirrors kv-cache-tester's default `--timing-strategy think-only`. Falls back to the full delta for turns whose recorded `think_time` is null. Mutually exclusive with ignore_trace_delays. No effect on non-weka trace loaders.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per Weka root trace. Only honored by format weka_trace: the loader drops whole traces whose *recorded* peak exceeds this ceiling at load time (filter-then-cap before entries). Not a DatasetManager tokenize filter; rejected for non-Weka formats.","title":"Maxcontextlength","x-jinja2-supported":true},"remoteFixedScheduleCheck":{"anyOf":[{"$ref":"#/$defs/RemoteFixedScheduleCheck"},{"type":"null"}],"default":null,"description":"Internal: set by the CLI for a remote `--input-file` and cleared by the dataset preflight once the first record decides fixed-schedule auto-promotion. Do not set in YAML."},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for trace replay (mirror of the synthetic dataset's prompts.cache_bust). The engine reads the active dataset's target via ``BenchmarkConfig.get_cache_bust_target()``; 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."}},"required":["name","type"],"title":"FileDataset","type":"object"},"FixedDistribution":{"title":"FixedDistribution","description":"Returns a constant value on every sample. Scalars coerce to this. Can be a number (interpreted as a fixed value) or {value} object.","anyOf":[{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable resolving to a fixed distribution value."},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template resolving to a fixed distribution value."},{"additionalProperties":false,"description":"Returns a constant value on every sample. Scalars coerce to this.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"value":{"description":"The constant value returned on every sample.","title":"Value","type":"number"},"type":{"const":"fixed","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["value"],"title":"FixedDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"integer"},{"type":"number"}]},"FixedSchedulePhase":{"additionalProperties":false,"description":"Replay requests at predetermined timestamps from a trace dataset.\n\nStop condition not required -- the trace dataset determines when the\nphase ends.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"fixed_schedule","description":"Replay requests at trace timestamps.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"autoOffset":{"default":true,"description":"Normalize trace timestamps to start at 0. Subtracts minimum timestamp from all entries.","title":"Autooffset","type":"boolean"},"startOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Filter out trace requests before this timestamp in ms (must be >= 0).","title":"Startoffset","x-jinja2-supported":true},"endOffset":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Filter out trace requests after this timestamp in ms (must be >= 0).","title":"Endoffset","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Fixed-schedule phases only accept disabled adaptive-scale no-op blocks.","title":"Adaptivescale","anyOf":[{"const":false},{"$ref":"#/$defs/DisabledAdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"FixedSchedulePhase","type":"object"},"GPUTelemetryCollectorType":{"enum":["dcgm","pynvml","amdsmi"],"title":"GPUTelemetryCollectorType","type":"string"},"GPUTelemetryMode":{"description":"GPU telemetry display mode.","enum":["summary","realtime_dashboard"],"title":"GPUTelemetryMode","type":"string"},"GammaPhase":{"additionalProperties":false,"description":"Gamma-distributed request arrivals with configurable smoothness.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"gamma","description":"Gamma-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"smoothness":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Gamma distribution shape parameter (must be > 0). 1.0 = Poisson, <1 = bursty, >1 = regular.","title":"Smoothness","x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"GammaPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"GpuTelemetryConfig":{"title":"GpuTelemetryConfig","description":"GPU telemetry configuration for live or replayed GPU metrics collection.\n\nCollects GPU metrics through DCGM exporter endpoints by default. The\n``collector`` field can switch to a local backend (``pynvml`` for NVIDIA,\n``amdsmi`` for AMD ROCm); ``mode`` controls summary vs. realtime dashboard\ndisplay.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9400/metrics\"\n → GpuTelemetryConfig(enabled=True, urls=[\"http://localhost:9400/metrics\"])\n - Singular url field: {url: \"...\"}\n → GpuTelemetryConfig(urls=[\"...\"]) Can be a URL string or full config object.","anyOf":[{"additionalProperties":false,"description":"GPU telemetry configuration for live or replayed GPU metrics collection.\n\nCollects GPU metrics through DCGM exporter endpoints by default. The\n``collector`` field can switch to a local backend (``pynvml`` for NVIDIA,\n``amdsmi`` for AMD ROCm); ``mode`` controls summary vs. realtime dashboard\ndisplay.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9400/metrics\"\n → GpuTelemetryConfig(enabled=True, urls=[\"http://localhost:9400/metrics\"])\n - Singular url field: {url: \"...\"}\n → GpuTelemetryConfig(urls=[\"...\"])","properties":{"enabled":{"default":true,"description":"Enable GPU telemetry collection. Set to false to disable.","title":"Enabled","type":"boolean"},"urls":{"description":"DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","items":{"type":"string"},"title":"Urls","type":"array"},"metricsFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to CSV file with pre-recorded GPU metrics. Alternative to live DCGM collection.","title":"Metricsfile"},"collector":{"$ref":"#/$defs/GPUTelemetryCollectorType","default":"dcgm","description":"GPU telemetry collector backend. Use 'dcgm' for DCGM exporter endpoints or a local collector (e.g. 'pynvml' for NVIDIA, 'amdsmi' for AMD ROCm) for on-host metrics collection."},"mode":{"$ref":"#/$defs/GPUTelemetryMode","default":"summary","description":"GPU telemetry display mode. Summary emits aggregate console output; realtime_dashboard enables live dashboard updates."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","oneOf":[{"type":"string"},{"description":"DCGM exporter endpoint URLs. Example: http://localhost:9400/metrics","items":{"type":"string"},"title":"Urls","type":"array"}],"x-singular-alias-of":"urls"}},"title":"GpuTelemetryConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"string"}]},"GridSweep":{"additionalProperties":false,"description":"Grid sweep - all combinations of parameters (Cartesian product).","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"grid","default":"grid","description":"Sweep type discriminator.","title":"Type","type":"string"},"parameters":{"additionalProperties":{"items":{},"type":"array"},"description":"Parameters to sweep: dot-notation path -> list of values.","minProperties":1,"title":"Parameters","type":"object"}},"required":["parameters"],"title":"GridSweep","type":"object"},"ImageConfig":{"additionalProperties":false,"allOf":[{"if":{"properties":{"source":{"const":"noise"}}},"then":{"properties":{"sourceSampling":{"const":"random-with-replacement"}}}}],"description":"Configuration for synthetic image generation in multimodal datasets.\n\nControls the generation of synthetic images for vision-language\nmodel benchmarking. Images are generated by randomly sampling and\nresizing source images to specified dimensions.","properties":{"batchSize":{"default":0,"description":"Number of images to include in each multimodal request. Supported with chat endpoint type for vision-language models. Set to 0 to disable image inputs. Higher batch sizes test multi-image understanding and increase request payload size.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of images to include in each multimodal request. Supported with chat endpoint type for vision-language models. Set to 0 to disable image inputs. Higher batch sizes test multi-image understanding and increase request payload size.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"width":{"description":"Image width in pixels. Can be a fixed integer or {mean, stddev} distribution. Combined with height to determine image dimensions and file sizes for multimodal benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Width","x-jinja2-supported":true},"height":{"description":"Image height in pixels. Can be a fixed integer or {mean, stddev} distribution. Used when batch_size > 0 for multimodal vision benchmarking.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Height","x-jinja2-supported":true},"format":{"$ref":"#/$defs/ImageFormat","default":"jpeg","description":"Image file format for generated images. png: lossless compression (larger files, best quality). jpeg: lossy compression (smaller files, good quality). random: randomly select between PNG and JPEG per image. Format affects file size in multimodal requests and encoding overhead."},"source":{"anyOf":[{"$ref":"#/$defs/ImageSource"},{"format":"path","type":"string"}],"default":"noise","description":"Source image generation mode (default: noise). noise: generate random noise images on the fly — no files on disk, effectively unbounded variety so servers cannot dedupe identical inputs. assets: index images from the bundled assets/source_images directory and lazily load them at the requested dimensions. A path string indexes images from the given directory (e.g. ./source_images). Random-noise images are roughly incompressible, so payload bytes are larger than equivalent natural images.","title":"Source"},"sourceSampling":{"$ref":"#/$defs/ImageSourceSamplingStrategy","default":"random-with-replacement","description":"How source images are selected from finite image sources selected by source='assets' or a directory path. random-with-replacement: draw each source image independently; repeats may occur immediately. shuffle-cycle: draw every source image once per shuffled cycle, reshuffling after exhaustion. sequential-cycle: walk source images in sorted load order and wrap after exhaustion. For noise mode, only random-with-replacement is valid because there is no finite source pool."}},"title":"ImageConfig","type":"object"},"ImageFormat":{"description":"Image file formats for synthetic image generation.","enum":["png","jpeg","random"],"title":"ImageFormat","type":"string"},"ImageSource":{"description":"Source image generation mode for multimodal benchmarking.","enum":["assets","noise"],"title":"ImageSource","type":"string"},"ImageSourceSamplingStrategy":{"description":"How source images are selected from a finite source-image pool.","enum":["random-with-replacement","shuffle-cycle","sequential-cycle"],"title":"ImageSourceSamplingStrategy","type":"string"},"IpcCommunicationConfig":{"additionalProperties":false,"description":"IPC (Unix socket) communication configuration.\n\nFor single-machine deployments with lowest latency.\nUses Unix domain sockets for all inter-service communication.","properties":{"type":{"const":"ipc","description":"Communication type. Must be 'ipc'.","title":"Type","type":"string"},"path":{"default":"/tmp/aiperf","description":"Directory for IPC socket files. AIPerf creates multiple socket files in this directory.","title":"Path","type":"string"}},"required":["type"],"title":"IpcCommunicationConfig","type":"object"},"JsonValue":{},"LatinHypercubeSweep":{"additionalProperties":false,"description":"Latin hypercube sampling — perfect marginal stratification per axis.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"samples":{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","title":"Samples","oneOf":[{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","maximum":1048576,"minimum":2,"title":"Samples","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"seed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"RNG seed for reproducibility. Must be non-negative; scipy's QMC engines reject negative seeds with an opaque low-level error.","title":"Seed","x-jinja2-supported":true},"dimensions":{"description":"Dimensions to sample over.","items":{"$ref":"#/$defs/SamplingDimension"},"minItems":1,"title":"Dimensions","type":"array"},"labelFormat":{"default":"index","description":"Variation-label format. 'index' => '_NNNN' (short, sortable). 'kv' => 'k1=v1, k2=v2' (grid-style).","enum":["index","kv"],"title":"Labelformat","type":"string"},"type":{"const":"latin_hypercube","default":"latin_hypercube","description":"Sweep type discriminator.","title":"Type","type":"string"},"optimization":{"default":"random-cd","description":"LHS post-processing to reduce centered discrepancy. None = vanilla LHS.","enum":[null,"random-cd","lloyd"],"title":"Optimization"}},"required":["samples","dimensions"],"title":"LatinHypercubeSweep","type":"object"},"LogNormalDistribution":{"additionalProperties":false,"description":"Log-normal parameterized by mean and median (right-skewed, always positive).\n\nSkew is controlled by the mean/median ratio: larger ratio = heavier right tail.\nWhen mean == median the distribution is deterministic.\n\nInternally: sigma = sqrt(2 * log(mean / median)), mu = log(median).","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"mean":{"description":"Desired mean of the output distribution.","title":"Mean","oneOf":[{"description":"Desired mean of the output distribution.","exclusiveMinimum":0.0,"title":"Mean","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"median":{"description":"Desired median. Must be <= mean. Lower median = more right skew.","title":"Median","oneOf":[{"description":"Desired median. Must be <= mean. Lower median = more right skew.","exclusiveMinimum":0.0,"title":"Median","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"type":{"const":"lognormal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["mean","median"],"title":"LogNormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"LoggingConfig":{"additionalProperties":false,"description":"Logging configuration for verbosity and debug settings.","properties":{"level":{"$ref":"#/$defs/AIPerfLogLevel","default":"INFO","description":"Global logging verbosity level (loguru-style severity ladder; TRACE most verbose, CRITICAL least). NOTICE and SUCCESS are loguru-specific intermediate levels (between INFO and WARNING) and have no equivalent in the Python stdlib `logging` module — pick INFO/WARNING/ERROR if you want stdlib-compatible severities."}},"title":"LoggingConfig","type":"object"},"MLflowConfig":{"additionalProperties":false,"description":"MLflow tracking and artifact-upload configuration.","properties":{"trackingUri":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"MLflow tracking URI.","title":"Trackinguri"},"experiment":{"default":"aiperf","description":"MLflow experiment name.","title":"Experiment","type":"string"},"runName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"MLflow run name.","title":"Runname"},"tags":{"anyOf":[{"type":"array","items":{"oneOf":[{"type":"string"},{"type":"array","minItems":2,"maxItems":2}]}},{"type":"string"},{"additionalProperties":{"type":"string"},"type":"object"},{"type":"null"}],"default":null,"description":"Additional MLflow run tags to attach on upload.","title":"Tags"},"parentRunId":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Optional MLflow parent run ID.","title":"Parentrunid"},"artifactGlobs":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Artifact glob overrides for MLflow upload.","title":"Artifactglobs"}},"title":"MLflowConfig","type":"object"},"MetricsConfig":{"additionalProperties":false,"description":"Configuration for benchmark metric aggregation behavior.\n\nCurrently empty — list-valued record metrics aggregate via t-digest at\nfirst-touch (see :mod:`aiperf.metrics.list_metric_aggregation`); the\nsketch compression is tunable via ``AIPERF_METRICS_TDIGEST_COMPRESSION``.\nReserved as the home for any future user-facing metric-aggregation knobs.","properties":{},"title":"MetricsConfig","type":"object"},"ModelItem":{"additionalProperties":false,"description":"Configuration for a single model in advanced models configuration.\n\nUsed when the models section uses the advanced format with\nexplicit items, weights, and per-model settings.","properties":{"name":{"description":"Model name or identifier as known to the inference server.","minLength":1,"title":"Name","type":"string"},"weight":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Selection weight for weighted strategy (0.0-1.0). Weights must sum to 1.0 (+/-0.01) across all models; they are validated, not auto-normalized. Example: weight=0.7 means ~70%% of requests to this model.","title":"Weight","x-jinja2-supported":true},"lora":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"LoRA adapter name to load with this model. Server must support dynamic LoRA adapter loading.","title":"Lora"},"modalities":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"List of input modalities this model supports. [Currently a no-op: no selection strategy consumes this field. Accepted for forward-compatibility / declarative documentation.]","title":"Modalities"},"tokenizer":{"anyOf":[{"$ref":"#/$defs/TokenizerOverride"},{"type":"null"}],"default":null,"description":"Per-model tokenizer override. Use when this model requires a different tokenizer than global config."}},"required":["name"],"title":"ModelItem","type":"object"},"ModelSelectionStrategy":{"description":"Strategy for selecting the model to use for the request.","enum":["round_robin","random","weighted"],"title":"ModelSelectionStrategy","type":"string"},"ModelsAdvanced":{"additionalProperties":false,"description":"Advanced models configuration with selection strategy and item details.\n\nUse this format when you need weighted routing, LoRA adapters,\nor per-model tokenizer overrides.","properties":{"strategy":{"$ref":"#/$defs/ModelSelectionStrategy","default":"round_robin","description":"Strategy for selecting models when multiple are configured. round_robin cycles through models, random selects randomly, weighted uses configured weights."},"items":{"description":"List of model configurations. At least one model required.","items":{"$ref":"#/$defs/ModelItem"},"minItems":1,"title":"Items","type":"array"}},"required":["items"],"title":"ModelsAdvanced","type":"object"},"MultiRunConfig":{"additionalProperties":false,"description":"Trial mechanics within a single sweep variation.\n\nWhen `num_runs > 1` (or `convergence` is set), AIPerf runs multiple\ntrials per variation and aggregates statistics. When `num_runs == 1`\nand `convergence is None`, exactly one trial runs per variation.","properties":{"numRuns":{"default":1,"description":"Upper bound on trials per variation. When `convergence` is set, trials may stop earlier; this is the hard ceiling. Cap matches `BenchmarkPlan.trials` so programmatic plan assembly cannot exceed what user-facing config validation allows.","title":"Numruns","oneOf":[{"default":1,"description":"Upper bound on trials per variation. When `convergence` is set, trials may stop earlier; this is the hard ceiling. Cap matches `BenchmarkPlan.trials` so programmatic plan assembly cannot exceed what user-facing config validation allows.","maximum":10,"minimum":1,"title":"Numruns","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"cooldownSeconds":{"default":0.0,"description":"Cooldown between trials within a single variation. Capped at 24h to surface typos like `1e18` at config-load time rather than blocking the orchestrator inside ``asyncio.sleep``.","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between trials within a single variation. Capped at 24h to surface typos like `1e18` at config-load time rather than blocking the orchestrator inside ``asyncio.sleep``.","maximum":86400,"minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"confidenceLevel":{"default":0.95,"description":"Confidence level for aggregate statistics (0.90, 0.95, 0.99).","title":"Confidencelevel","oneOf":[{"default":0.95,"description":"Confidence level for aggregate statistics (0.90, 0.95, 0.99).","exclusiveMaximum":1,"exclusiveMinimum":0,"title":"Confidencelevel","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"setConsistentSeed":{"default":true,"description":"Auto-set a random seed when none is given for cross-trial consistency.","title":"Setconsistentseed","type":"boolean"},"varySeedPerTrial":{"default":false,"description":"When True, derive a distinct seed for each trial of a variation via SHA-256 over `(envelope_seed, variation.label, trial)`. When False (default), all trials of a variation share the variation's seed. Off gives pure-runtime variance for confidence statistics; on captures end-to-end variance including per-trial workload sampling.","title":"Varyseedpertrial","type":"boolean"},"disableWarmupAfterFirst":{"default":true,"description":"Skip warmup on trials 2..N of a variation (steady-state measurement).","title":"Disablewarmupafterfirst","type":"boolean"},"convergence":{"anyOf":[{"$ref":"#/$defs/ConvergenceConfig"},{"type":"null"}],"default":null,"description":"Optional adaptive early-stop. Absent = fixed-N (run `num_runs` trials)."}},"title":"MultiRunConfig","type":"object"},"MultimodalDistribution":{"additionalProperties":false,"description":"Weighted mixture of N peaks (N >= 2).\n\nYAML:\n isl:\n peaks:\n - {mean: 128, stddev: 20, weight: 60}\n - {mean: 2048, median: 1800, weight: 40}\n # Equal split — omit weight:\n isl:\n peaks:\n - {mean: 128, stddev: 20}\n - {mean: 2048, median: 1800}\n - {mean: 8192, median: 4096}","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"peaks":{"description":"Two or more weighted sub-distributions.","items":{"$ref":"#/$defs/PeakEntry"},"minItems":2,"title":"Peaks","type":"array"},"type":{"const":"multimodal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["peaks"],"title":"MultimodalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"NetworkLatencyConfig":{"additionalProperties":false,"description":"Network latency calibration configuration.\n\nWhen ``enabled``, AIPerf opens a fresh TCP connection to the endpoint\nhost:port on an interval during profiling and records the handshake RTT.\nThe mean RTT is subtracted from the request-start-anchored latency metrics\n(request_latency, time_to_first_token, time_to_first_output_token) to produce\nnon-destructive ``network_adjusted_*``\nvariants. ``mean_ms`` supplies a fixed mean RTT and skips active probing.","properties":{"enabled":{"default":false,"description":"Enable network latency calibration. When true, AIPerf measures the endpoint RTT (or uses mean_ms) and subtracts it from latency metrics, emitted as separate network_adjusted_* metrics.","title":"Enabled","type":"boolean"},"pingInterval":{"description":"Seconds between TCP-handshake RTT probes during profiling. Ignored when mean_ms is set.","title":"Pinginterval","oneOf":[{"description":"Seconds between TCP-handshake RTT probes during profiling. Ignored when mean_ms is set.","exclusiveMinimum":0.0,"title":"Pinginterval","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"meanMs":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Fixed mean network RTT in milliseconds to subtract, bypassing active probing. When set, no probes are sent and this value is used directly.","title":"Meanms","x-jinja2-supported":true}},"title":"NetworkLatencyConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},"NormalDistribution":{"additionalProperties":false,"description":"Gaussian (truncated at 0) parameterized by mean and stddev.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min","x-jinja2-supported":true},"max":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max","x-jinja2-supported":true},"mean":{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","title":"Mean","oneOf":[{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","minimum":0.0,"title":"Mean","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"stddev":{"default":0.0,"description":"Standard deviation. 0 = deterministic.","title":"Stddev","oneOf":[{"default":0.0,"description":"Standard deviation. 0 = deterministic.","minimum":0.0,"title":"Stddev","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"type":{"const":"normal","description":"Optional distribution type marker; stripped before concrete validation."}},"required":["mean"],"title":"NormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},"OTelConfig":{"additionalProperties":false,"description":"OpenTelemetry metrics streaming configuration.","properties":{"metricsUrl":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"OTLP/HTTP metrics endpoint URL.","title":"Metricsurl"},"streamMetricsEnabled":{"default":true,"description":"Stream metric records to OTel.","title":"Streammetricsenabled","type":"boolean"},"streamTimingEnabled":{"default":true,"description":"Stream timing records to OTel.","title":"Streamtimingenabled","type":"boolean"},"customResourceAttributes":{"additionalProperties":{"type":"string"},"description":"Custom OTel resource attributes.","title":"Customresourceattributes","type":"object"},"genAiProvider":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"GenAI semantic convention provider override.","title":"Genaiprovider"}},"title":"OTelConfig","type":"object"},"Objective":{"additionalProperties":false,"description":"Single optimization objective.\n\nMultiple ``Objective`` entries on an ``AdaptiveSearchSweep`` enable\nPareto Bayesian optimization (e.g. qlognehvi / qnehvi / qehvi). A\nlength-1 list is single-objective BO.","properties":{"metric":{"description":"Metric tag to optimize (e.g. 'output_token_throughput').","title":"Metric","type":"string"},"stat":{"default":"avg","description":"Statistic on the metric.","enum":["avg","p50","p90","p95","p99"],"title":"Stat","type":"string"},"direction":{"$ref":"#/$defs/OptimizationDirection","description":"MAXIMIZE or MINIMIZE."},"threshold":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Pareto reference point for hypervolume computation. Trials worse than this on this objective do not contribute to hypervolume. When None, auto-derived from the worst observed value among Sobol initial points. Ignored for single-objective runs.","title":"Threshold","x-jinja2-supported":true}},"required":["metric","direction"],"title":"Objective","type":"object"},"OptimizationDirection":{"description":"Direction of optimization for a metric.\n\nMAXIMIZE: Higher values are preferred (e.g. throughput, goodput).\nMINIMIZE: Lower values are preferred (e.g. latency, TTFT, p99).","enum":["maximize","minimize"],"title":"OptimizationDirection","type":"string"},"OutcomeConstraint":{"additionalProperties":false,"description":"Feasibility gate on a metric the optimizer is *not* optimizing.\n\nDistinct from ``Objective.threshold`` (Pareto reference point) and from\n``SLAFilter`` (post-hoc filter on benchmark eligibility). An outcome\nconstraint masks BoTorch acquisition: candidates predicted to violate\nit are downweighted.","properties":{"metric":{"description":"Metric tag to constrain.","title":"Metric","type":"string"},"op":{"description":"Comparison operator.","enum":["<=",">=","=="],"title":"Op","type":"string"},"bound":{"description":"Threshold value.","title":"Bound","oneOf":[{"description":"Threshold value.","title":"Bound","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metric","op","bound"],"title":"OutcomeConstraint","type":"object"},"PeakEntry":{"title":"PeakEntry","description":"A weighted component in a multimodal distribution.\n\nThe weight and distribution fields are written inline in YAML:\n {mean: 128, stddev: 20, weight: 60}\n\nThe ``weight`` key is extracted before the remaining fields are parsed\nas a SamplingDistribution. Defaults to 1.0 (equal split when omitted).","anyOf":[{"additionalProperties":false,"description":"A weighted component in a multimodal distribution.\n\nThe weight and distribution fields are written inline in YAML:\n {mean: 128, stddev: 20, weight: 60}\n\nThe ``weight`` key is extracted before the remaining fields are parsed\nas a SamplingDistribution. Defaults to 1.0 (equal split when omitted).","properties":{"distribution":{"description":"The sub-distribution for this peak.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}],"title":"Distribution"},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["distribution"],"title":"PeakEntry","type":"object"},{"additionalProperties":false,"description":"Inline FixedDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"value":{"description":"The constant value returned on every sample.","title":"Value","type":"number"},"type":{"const":"fixed","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["value"],"title":"FixedDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline NormalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"mean":{"description":"Mean value. Must be >= 0; samples below 0 are truncated, so a negative mean would yield a degenerate distribution. Zero is allowed (e.g. OSL=0 disables output, turn_delay mean=0 disables inter-turn delay).","minimum":0.0,"title":"Mean","type":"number"},"stddev":{"default":0.0,"description":"Standard deviation. 0 = deterministic.","minimum":0.0,"title":"Stddev","type":"number"},"type":{"const":"normal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["mean"],"title":"NormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline LogNormalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"mean":{"description":"Desired mean of the output distribution.","exclusiveMinimum":0.0,"title":"Mean","type":"number"},"median":{"description":"Desired median. Must be <= mean. Lower median = more right skew.","exclusiveMinimum":0.0,"title":"Median","type":"number"},"type":{"const":"lognormal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["mean","median"],"title":"LogNormalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline MultimodalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"peaks":{"description":"Two or more weighted sub-distributions.","items":{"$ref":"#/$defs/PeakEntry"},"minItems":2,"title":"Peaks","type":"array"},"type":{"const":"multimodal","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["peaks"],"title":"MultimodalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true},{"additionalProperties":false,"description":"Inline EmpiricalDistribution peak with optional weight.","properties":{"min":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive lower bound; samples below are clamped up. Applies to every distribution type — composes with mean/stddev/median/peaks/points/value.","title":"Min"},"max":{"anyOf":[{"type":"number"},{"type":"null"}],"default":null,"description":"Inclusive upper bound; samples above are clamped down.","title":"Max"},"points":{"description":"Weighted discrete values. Weights are relative.","items":{"$ref":"#/$defs/EmpiricalPoint"},"title":"Points","type":"array"},"type":{"const":"empirical","description":"Optional distribution type marker; stripped before concrete validation."},"weight":{"default":1.0,"description":"Relative weight (normalised internally).","minimum":0.0,"title":"Weight","type":"number"}},"required":["points"],"title":"EmpiricalDistribution","type":"object","x-kubernetes-preserve-unknown-fields":true}]},"PoissonPhase":{"additionalProperties":false,"description":"Poisson-distributed request arrivals at the target rate.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"poisson","description":"Poisson-distributed rate-controlled arrivals.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"rateSeries":{"anyOf":[{"$ref":"#/$defs/RateSeriesConfig"},{"type":"null"}],"default":null,"description":"Piecewise-linear request-rate schedule."},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type"],"title":"PoissonPhase","type":"object","allOf":[{"oneOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}},{"required":["rateSeries"],"properties":{"rateSeries":{"not":{"type":"null"}}}}]}]},"PostProcessSpec":{"additionalProperties":false,"description":"Post-aggregation hook spec emitted by a search recipe.\n\nResolved through the ``search_recipe_post_process`` plugin category and\ninvoked by ``aggregate_sweep_and_export`` after per-variation aggregation.\n\nExample: ``PostProcessSpec(handler=\"ttft_sla_curve\", params={\"sla_ms\": 200},\noutput_filename=\"ttft_sla_curve.json\")``.","properties":{"handler":{"description":"Registered post-process handler name. Looked up in the ``search_recipe_post_process`` plugin category.","title":"Handler","type":"string"},"params":{"additionalProperties":true,"description":"Free-form parameters forwarded to the handler. Shape is handler-defined; Any is intentional because handlers ingest dynamic recipe-derived inputs.","title":"Params","type":"object"},"output_filename":{"description":"Filename (relative to ``sweep_aggregate/`` under the artifact dir) to write the handler's output to. Must end in ``.json`` and contain no path separators -- the hook joins it with ``sweep_aggregate/`` and refuses to escape that directory. NUL bytes and dot-only stems are also rejected.","title":"Output Filename","type":"string"}},"required":["handler","output_filename"],"title":"PostProcessSpec","type":"object"},"PrefixPromptConfig":{"additionalProperties":false,"description":"Configuration for prefix prompts (KV cache testing).\n\nPrefix prompts allow testing KV cache efficiency by generating\nrequests that share common prefixes. This simulates scenarios\nlike system prompts or shared context that can be cached.\n\nNote: pool_size/length are mutually exclusive with shared_system_length\nand user_context_length.","properties":{"poolSize":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Number of distinct prefix prompts to generate for KV cache testing. Each prefix is prepended to user prompts, simulating cached context scenarios. Prefixes are randomly selected from pool per request. Mutually exclusive with shared_system_length/user_context_length.","title":"Poolsize","x-jinja2-supported":true},"length":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Token length for each prefix prompt in the pool. Only used when pool_size is set. The prefix is concatenated to the body at the token level, so the final prompt contains exactly prefix length + input length tokens. Mutually exclusive with shared_system_length/user_context_length.","title":"Length","x-jinja2-supported":true},"sharedSystemLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Length of shared system prompt in tokens. This prompt is identical across all sessions and appears as a system message. First part of a two-part prefix structure with high cache hit rate expected. Mutually exclusive with pool_size/length.","title":"Sharedsystemlength","x-jinja2-supported":true},"userContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Length of per-session user context prompt in tokens. Each dataset entry gets a unique user context prompt. Second part of two-part prefix structure with lower cache hit rate expected. Mutually exclusive with pool_size/length.","title":"Usercontextlength","x-jinja2-supported":true}},"title":"PrefixPromptConfig","type":"object"},"PromptConfig":{"additionalProperties":false,"description":"Configuration for prompt/token specifications in synthetic datasets.\n\nThis is the core configuration for controlling input sequence length (ISL)\nand output sequence length (OSL) in synthetic data generation.","properties":{"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Input sequence length in tokens. Can be a fixed integer (e.g., 512) or distribution {mean: 512, stddev: 50}. AIPerf generates prompts with lengths following a normal distribution around the mean (±stddev). Ignored when sequence_distribution is specified.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length (max tokens to request via max_completion_tokens). Can be a fixed integer or distribution {mean, stddev}. Controls response length for synthetic datasets. When not set, the model determines output length. Ignored when sequence_distribution is specified.","title":"Osl"},"blockSize":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Token block size for hash-based prompt caching in mooncake_trace datasets. When hash_ids are provided in trace entries, prompts are divided into blocks of this size. Each hash_id maps to a cached block, enabling simulation of KV-cache sharing patterns from production workloads. Total prompt length = (num_hash_ids - 1) * block_size + final_block_size.","title":"Blocksize","x-jinja2-supported":true},"batchSize":{"default":1,"description":"Number of text inputs to include in each request for batch processing endpoints. Supported by embeddings and rankings endpoint types where models can process multiple inputs simultaneously. Set to 1 for single-input requests. Not applicable to chat or completions endpoints.","title":"Batchsize","oneOf":[{"default":1,"description":"Number of text inputs to include in each request for batch processing endpoints. Supported by embeddings and rankings endpoint types where models can process multiple inputs simultaneously. Set to 1 for single-input requests. Not applicable to chat or completions endpoints.","minimum":1,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker configuration. Injects a deterministic marker into prompts to defeat KV-cache prefix matching. Disabled by default (target='none')."},"corpus":{"anyOf":[{"$ref":"#/$defs/PromptCorpus"},{"type":"null"}],"default":null,"description":"Source corpus for synthetic prompt text generation. 'sonnet' uses Shakespeare sonnets. 'coding' uses realistic coding content (code, bash output, JSON, error tracebacks, git diffs). When unset, the active dataset loader's default applies (most loaders default to 'sonnet'; agentic-coding loaders such as weka_trace default to 'coding')."},"sequenceDistribution":{"anyOf":[{"items":{"$ref":"#/$defs/SequenceDistributionEntry"},"type":"array"},{"type":"null"}],"default":null,"description":"Distribution of (ISL, OSL) pairs with probabilities for mixed workload simulation. Each entry specifies {isl, osl, probability}. Probabilities are percentages (0-100) and must sum to 100. When specified, requests are sampled from this distribution instead of using isl/osl fields.","title":"Sequencedistribution"},"randomRangeRatio":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Sample ISL and OSL uniformly from a ratio-defined integer window around the configured means. Accepts a single float (applied to both ISL and OSL) or a JSON object {\"input\": 0.3, \"output\": 0.5} for independent values. Mutually exclusive with sequence_distribution. When a tokenizer is configured, the ISL mean is automatically reduced by tokenizer.num_special_tokens_to_add(pair=False) so --isl represents total server-side input tokens.","title":"Randomrangeratio"},"randomCorpusStyle":{"$ref":"#/$defs/RandomCorpusStyle","default":"vllm","description":"Sampling formula for random_range_ratio. vllm (default): symmetric window [floor(mean*(1-r)), ceil(mean*(1+r))], ratio in [0, 1). sglang: lower-bounded window [max(1, int(mean*r)), mean], ratio in [0, 1]."}},"title":"PromptConfig","type":"object"},"PromptCorpus":{"description":"Corpus used for synthetic prompt text generation.","enum":["sonnet","coding","random"],"title":"PromptCorpus","type":"string"},"PromptSelectionConfig":{"additionalProperties":false,"description":"Slim prompts block for file/public datasets (corpus selection only).","properties":{"corpus":{"anyOf":[{"$ref":"#/$defs/PromptCorpus"},{"type":"null"}],"default":null,"description":"Source corpus for synthesized prompt text. 'sonnet' uses Shakespeare sonnets. 'coding' uses realistic coding content. When unset, the active dataset loader's default applies. Honored only where content is synthesized (synthetic + hash/trace replay); verbatim loaders ignore it."}},"title":"PromptSelectionConfig","type":"object"},"PublicDataset":{"additionalProperties":false,"description":"Public dataset configuration.\n\nUses well-known public benchmarking datasets that are\nautomatically downloaded and processed by AIPerf.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"public","description":"Dataset type discriminator. Must be 'public'.","title":"Type","type":"string"},"dataset":{"$ref":"#/$defs/PublicDatasetType","description":"Pre-configured public dataset to download and use for benchmarking. Name of the HuggingFace public dataset enum (e.g. 'sharegpt', 'alpaca'). AIPerf automatically downloads and parses these datasets."},"entries":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Limit number of records to use from the dataset. If not specified, uses all available records.","title":"Entries","x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic sampling from the dataset. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"hfSubset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset subset/config name override (e.g. 'sharegpt4o'). Only applies for HuggingFace-backed public dataset loaders. Takes priority over the subset defined in the plugin registry.","title":"Hfsubset"},"filters":{"additionalProperties":{"type":"string"},"description":"Dataset-specific filters forwarded to public dataset loaders. Supported keys and values depend on the selected dataset.","title":"Filters","type":"object"},"hfMooncakeDataset":{"anyOf":[{"minLength":1,"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace repository for mooncake_hf; required for this loader. The CLI flag --hf-mooncake-dataset auto-selects this loader.","title":"Hfmooncakedataset"},"hfMooncakeSplit":{"anyOf":[{"minLength":1,"pattern":"\\S","type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace split override for mooncake_hf (e.g. train or test). Unset uses the plugin default, test. Only supported by mooncake_hf.","title":"Hfmooncakesplit"},"mooncakeTraceToHfMapping":{"additionalProperties":{"type":"string"},"description":"Mooncake target field to dotted source column path for mooncake_hf. Supports messages, session_id, tools, output_length, extra, and extra.PARAMETER.","title":"Mooncaketracetohfmapping","type":"object"},"hfWekaDataset":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace dataset repo for the generic Weka loader (e.g. `semianalysisai/cc-traces-weka-061526`). Only applies with `dataset: weka_hf`; setting it with any other public dataset is an error. Passing `--hf-weka-dataset` on the CLI auto-selects `--public-dataset weka_hf`, so the repo flag works on its own. Pinned Weka public dataset aliases keep their registry-defined repo names.","title":"Hfwekadataset"},"wekaNestedTimestampBasis":{"anyOf":[{"enum":["auto","absolute","relative"],"type":"string"},{"type":"null"}],"default":null,"description":"Corpus-wide interpretation of timestamps nested inside Weka subagent markers. Unset uses auto for Weka inputs. Auto selects relative if any child precedes its marker by more than 1 microsecond, otherwise absolute; this heuristic cannot certify uniform conventions. Absolute uses root-trace time and relative adds the marker time.","title":"Wekanestedtimestampbasis"},"interTurnDelayCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Clamp per-turn replay delays to at most this many seconds for HF-backed Weka trace replay (mirror of ``FileDataset.inter_turn_delay_cap_seconds``). ``None`` disables the cap.","title":"Interturndelaycapseconds","x-jinja2-supported":true},"traceIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Hard ceiling (seconds) for observed runtime idle time within each AgentX trajectory tree for HF-backed Weka replay (mirror of ``FileDataset.trace_idle_gap_cap_seconds``). Dataset timestamps remain unchanged. Defaults to None (no per-trace runtime idle cap).","title":"Traceidlegapcapseconds","x-jinja2-supported":true},"ignoreTraceDelays":{"default":false,"description":"Strip per-turn timestamps and inter-turn delays from HF-backed Weka trace replay at load time (mirror of ``FileDataset.ignore_trace_delays``). Mutually exclusive with use_think_time_only.","title":"Ignoretracedelays","type":"boolean"},"useThinkTimeOnly":{"default":false,"description":"For HF-backed Weka replay, emit Turn.delay using only the recorded per-request ``think_time`` instead of the full inter-request delta (mirror of ``FileDataset.use_think_time_only``). Mutually exclusive with ignore_trace_delays.","title":"Usethinktimeonly","type":"boolean"},"maxContextLength":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum peak prompt+output context length (tokens) per conversation for HF-backed Weka replay; drops over-length traces at load using *recorded* lengths (mirror of ``FileDataset.max_context_length``). Rejected for non-Weka public datasets.","title":"Maxcontextlength","x-jinja2-supported":true},"cacheBust":{"$ref":"#/$defs/CacheBustConfig","description":"Per-conversation cache-bust marker injection for HF-backed Weka trace replay (mirror of ``FileDataset.cache_bust``). 'none' (default) disables it."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptSelectionConfig"},{"type":"null"}],"default":null,"description":"Prompt synthesis selection for this dataset. Set ``corpus`` to choose sonnet vs coding when content is synthesized (trace hash_id reconstruction). Verbatim formats ignore it."},"synthesis":{"anyOf":[{"$ref":"#/$defs/SynthesisConfig"},{"type":"null"}],"default":null,"description":"Trace synthesis/transformation configuration for HF-backed Weka trace replay (mirror of ``FileDataset.synthesis``). ``--max-isl``/``--max-osl`` cap the per-conversation input/output lengths of the replayed HF Weka traces; without this they applied only to file-based traces. ``None`` disables synthesis."},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Output sequence length to apply when records do not specify one (mirror of ``FileDataset.osl``). Can be a fixed integer or {mean, stddev} distribution. Per-record output lengths always take precedence.","title":"Osl"}},"required":["name","type","dataset"],"title":"PublicDataset","type":"object"},"PublicDatasetType":{"enum":["exgentic_v2","exgentic","sharegpt","aimo","mmstar","mmvu","vision_arena","llava_onevision","aimo_aime","aimo_numina_cot","aimo_numina_1_5","spec_bench","spec_al_gsm8k","spec_al_math500","spec_al_humaneval","spec_al_mbpp","spec_al_mtbench","instruct_coder","blazedit_5k","blazedit_10k","librispeech","voxpopuli","gigaspeech","ami","spgispeech","semianalysis_cc_traces_weka","semianalysis_cc_traces_weka_no_subagents","semianalysis_cc_traces_weka_with_subagents","semianalysis_cc_traces_weka_with_subagents_256k","semianalysis_cc_traces_weka_with_subagents_060226","semianalysis_cc_traces_weka_with_subagents_060226_256k","semianalysis_cc_traces_weka_with_subagents_060526","semianalysis_cc_traces_weka_with_subagents_060526_256k","semianalysis_cc_traces_weka_with_subagents_060826","semianalysis_cc_traces_weka_with_subagents_060826_256k","semianalysis_cc_traces_weka_061326","semianalysis_cc_traces_weka_061326_256k","semianalysis_cc_traces_weka_061526","semianalysis_cc_traces_weka_061526_256k","semianalysis_cc_traces_weka_062126","semianalysis_cc_traces_weka_062126_256k","weka_hf","mooncake_hf","speed_bench_qualitative","speed_bench_coding","speed_bench_humanities","speed_bench_math","speed_bench_multilingual","speed_bench_qa","speed_bench_rag","speed_bench_reasoning","speed_bench_roleplay","speed_bench_stem","speed_bench_summarization","speed_bench_writing","speed_bench_throughput_1k","speed_bench_throughput_1k_low_entropy","speed_bench_throughput_1k_mixed","speed_bench_throughput_1k_high_entropy","speed_bench_throughput_2k","speed_bench_throughput_2k_low_entropy","speed_bench_throughput_2k_mixed","speed_bench_throughput_2k_high_entropy","speed_bench_throughput_8k","speed_bench_throughput_8k_low_entropy","speed_bench_throughput_8k_mixed","speed_bench_throughput_8k_high_entropy","speed_bench_throughput_16k","speed_bench_throughput_16k_low_entropy","speed_bench_throughput_16k_mixed","speed_bench_throughput_16k_high_entropy","speed_bench_throughput_32k","speed_bench_throughput_32k_low_entropy","speed_bench_throughput_32k_mixed","speed_bench_throughput_32k_high_entropy","speed_bench_throughput_100k","speed_bench_throughput_100k_low_entropy","speed_bench_throughput_100k_mixed","speed_bench_throughput_100k_high_entropy","h_cua_perf"],"title":"PublicDatasetType","type":"string"},"RampConfig":{"title":"RampConfig","description":"Configuration for gradual value ramping.\n\nControls how a value (concurrency, rate, etc.) transitions from\nstart to target over time. Can be a duration (number in seconds or string like '30s', '5m', '2h') or {duration, strategy} object.","anyOf":[{"additionalProperties":false,"description":"Configuration for gradual value ramping.\n\nControls how a value (concurrency, rate, etc.) transitions from\nstart to target over time.","properties":{"duration":{"description":"Seconds to ramp from start to target value.","exclusiveMinimum":0.0,"title":"Duration","type":"number"},"strategy":{"$ref":"#/$defs/RampType","default":"linear","description":"Ramp curve shape: linear (constant rate), exponential (slow start, fast finish), poisson (stochastic with guaranteed completion)."}},"required":["duration"],"title":"RampConfig","type":"object"},{"type":"integer"},{"type":"number"},{"type":"string"}]},"RampType":{"enum":["exponential","linear","poisson"],"title":"RampType","type":"string"},"RandomCorpusStyle":{"description":"Benchmark style for RANDOM corpus generation.\n\nControls the full set of behaviors that vary between tools. Each style\nis a bundle of decisions that together reproduce the statistical\ndistribution of a specific benchmarking tool:\n\n.. list-table::\n :header-rows: 1\n :widths: 30 35 35\n\n * - Concern\n - VLLM\n - SGLANG\n * - Token pool\n - ``valid_token_ids`` (special tokens excluded via ``all_special_ids``)\n - ``all_token_ids`` (full ``range(vocab_size)``, no exclusion)\n * - BOS / special-token adjustment\n - ``max(0, mean - num_special_tokens)``, applied to the mean before\n the window bounds are computed\n - ``max(1, drawn - num_special_tokens)``, applied per-request after\n sampling so the window keeps its raw-mean shape\n * - ISL/OSL range formula\n - Symmetric: ``[floor(mean*(1-r)), ceil(mean*(1+r))]``\n - Lower-bounded: ``[max(1, int(mean*r)), mean]``\n * - RNG algorithm (preseed)\n - PCG64 via ``numpy.random.default_rng(seed)``\n - MT19937 via a private ``numpy.random.RandomState(seed)``\n * - RNG draw order (with preseed)\n - All ISLs → all OSLs → all offsets\n - All ISLs → all OSLs → all offsets (same order, different algorithm)\n * - Top-up RNG\n - Continues from the same ``default_rng`` stream (``_preseed_rng``)\n - Continues from the same ``RandomState`` stream (``_preseed_rng``)\n\nBoth styles are preseeded whenever the corpus is\n:attr:`PromptCorpus.RANDOM` — the composer gates on\n``isinstance(dist, RangeRatioDistribution)``, which the SGLANG subclass\nsatisfies. ``_corpus_rng`` is the fallback for the non-preseeded paths\n(non-RANDOM corpora, and draws past cache exhaustion), not a per-style\ndifference.\n\n.. note::\n The two styles subtract special tokens at different points on purpose,\n each following its own upstream. VLLM folds the count into the bounds;\n SGLANG shifts each drawn length, mirroring ``sample_random_requests``\n under ``return_text`` (the default, and what aiperf sends)::\n\n input_lens[i] = max(1, input_lens[i] - num_special_tokens)\n\n Both land the same wire ISL on the configured value; only the resulting\n window shape differs.","enum":["vllm","sglang"],"title":"RandomCorpusStyle","type":"string"},"RankingsConfig":{"additionalProperties":false,"description":"Configuration for rankings/reranking endpoint datasets.\n\nControls the generation of query-passage pairs for benchmarking\nreranking and ranking models. Each request contains one query\nand multiple passages to rank.","properties":{"passages":{"description":"Number of passages per ranking request. Can be a fixed integer or {mean, stddev} distribution. Higher values test ranking at scale but increase request payload size and processing time.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Passages","x-jinja2-supported":true},"passageTokens":{"description":"Token length for each passage in ranking requests. Can be a fixed integer or {mean, stddev} distribution. Passages are synthetically generated text. Longer passages increase input processing demands and request size.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Passagetokens","x-jinja2-supported":true},"queryTokens":{"description":"Token length for the query text in ranking requests. Can be a fixed integer or {mean, stddev} distribution. Each ranking request contains one query and multiple passages.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Querytokens","x-jinja2-supported":true}},"title":"RankingsConfig","type":"object"},"RateSeriesConfig":{"additionalProperties":false,"description":"Piecewise-linear request-rate schedule.","properties":{"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"JSON file path containing request-rate control points.","title":"Path"},"points":{"description":"Strictly increasing request-rate control points.","items":{"$ref":"#/$defs/RateSeriesPoint"},"title":"Points","type":"array","minItems":2}},"title":"RateSeriesConfig","type":"object","allOf":[{"oneOf":[{"required":["path"],"properties":{"path":{"not":{"type":"null"}}}},{"required":["points"],"properties":{"points":{"minItems":2}}}]}]},"RateSeriesPoint":{"additionalProperties":false,"description":"One request-rate control point.","properties":{"timeS":{"description":"Elapsed time in seconds from the start of the request-rate series, after any request-rate ramp completes.","title":"Times","oneOf":[{"description":"Elapsed time in seconds from the start of the request-rate series, after any request-rate ramp completes.","minimum":0.0,"title":"Times","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"qps":{"description":"Request rate in queries per second at this point.","title":"Qps","oneOf":[{"description":"Request rate in queries per second at this point.","exclusiveMinimum":0.0,"title":"Qps","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["timeS","qps"],"title":"RateSeriesPoint","type":"object"},"RemoteFixedScheduleCheck":{"additionalProperties":false,"description":"Fixed-schedule auto-promotion deferred until a remote trace's first record is readable.\n\nSet by the CLI converter for a remote --input-file; consumed and cleared by the\ndataset access preflight.","properties":{"promoted":{"description":"True when the CLI built the profiling phase as fixed_schedule ahead of the timestamp check; the access preflight demotes it if the trace has none.","title":"Promoted","type":"boolean"},"conflictError":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Error raised if the trace turns out to have timestamps (an incompatible rate-shaped mode or a parameter sweep was selected).","title":"Conflicterror"},"demotedRequests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Request count the CLI would have defaulted to without the promotion (scaled to concurrency); applied if the access preflight demotes the phase. None when the user set a stop condition.","title":"Demotedrequests","x-jinja2-supported":true},"demotedStopWarning":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Warning logged when demoted_requests is applied.","title":"Demotedstopwarning"}},"required":["promoted"],"title":"RemoteFixedScheduleCheck","type":"object"},"RemoteRef":{"anyOf":[{"type":"string"},{"additionalProperties":{"type":"string"},"type":"object"}]},"RequestContentType":{"description":"Content type for HTTP request body serialization.","enum":["application/json","multipart/form-data"],"title":"RequestContentType","type":"string"},"RequestSignerType":{"enum":["sigv4"],"title":"RequestSignerType","type":"string"},"ResetKvCacheConfig":{"additionalProperties":false,"properties":{"timeoutSeconds":{"anyOf":[{"gt":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Timeout in seconds for the reset_kv_cache POST. When unset, uses 30.0s.","title":"Timeoutseconds","x-jinja2-supported":true},"maxRetrySeconds":{"anyOf":[{"ge":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total time budget in seconds for retrying a retryable reset_kv_cache POST failure: a timeout/connection error, or a response with status in RESET_KV_CACHE_RETRYABLE_STATUS_CODES (409/423/429/503). Other non-2xx responses are never retried. When unset, uses 60.0s.","title":"Maxretryseconds","x-jinja2-supported":true},"path":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for KV-cache reset. Default: /reset_prefix_cache.","title":"Path"}},"title":"ResetKvCacheConfig","type":"object"},"RuntimeConfig":{"additionalProperties":false,"description":"Runtime configuration for benchmark execution.","properties":{"ui":{"$ref":"#/$defs/UIType","default":"dashboard","description":"User interface mode. dashboard: rich interactive UI, simple: text progress, none: silent operation."},"workers":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total worker processes across the deployment (the maximum, ramped up to from `workers_min`). null = auto-detect based on CPU cores. Distinct from `workers_per_pod` (per-Kubernetes-pod fan-out) and `workers_min` (lower bound for adaptive ramp).","title":"Workers","x-jinja2-supported":true},"recordProcessors":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Total parallel record processors across the deployment. null = auto-detect based on CPU cores. Distinct from `record_processors_per_pod` (per-Kubernetes-pod fan-out).","title":"Recordprocessors","x-jinja2-supported":true},"serviceRunType":{"$ref":"#/$defs/ServiceRunType","default":"multiprocessing","description":"Execution mode. multiprocessing: local multi-process (default for `aiperf profile`). kubernetes: distributed across pods. [operator-managed under AIPerfJob — the operator forces 'kubernetes' on the controller pod; do not set this in a CR spec.]"},"communication":{"anyOf":[{"discriminator":{"propertyName":"type","mapping":{"ipc":"#/$defs/IpcCommunicationConfig","tcp":"#/$defs/TcpCommunicationConfig","dual":"#/$defs/DualBindCommunicationConfig"}},"anyOf":[{"$ref":"#/$defs/IpcCommunicationConfig"},{"$ref":"#/$defs/TcpCommunicationConfig"},{"$ref":"#/$defs/DualBindCommunicationConfig"}]},{"type":"null"}],"default":null,"description":"Inter-process communication configuration. Defaults to IPC for single-machine operation.","title":"Communication"},"apiPort":{"anyOf":[{"maximum":65535,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AIPerf API server port. Enables HTTP and WebSocket endpoints for real-time metrics and control.","title":"Apiport","x-jinja2-supported":true},"apiHost":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"AIPerf API server host. Requires api_port to be set.","title":"Apihost"},"datasetApiBaseUrl":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Base URL the operator injects into worker pods so they can fetch datasets from the controller-pod sidecar in Kubernetes mode. [operator-managed; do not set in a CR spec.]","title":"Datasetapibaseurl"},"workersPerPod":{"anyOf":[{"maximum":100,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Worker containers packed into each Kubernetes worker pod (Kubernetes mode only). The total number of worker containers cluster-wide is `workers` (or auto-detected); this knob controls how that total is fanned across pods. Ignored in multiprocessing mode. Sizing guide: each worker was designed around ~500 concurrent connections, and a node's ~65k ephemeral ports is the ceiling this fan-out exists to work around.","title":"Workersperpod","x-jinja2-supported":true},"recordProcessorsPerPod":{"anyOf":[{"maximum":100,"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Record-processor containers packed into each Kubernetes worker pod (Kubernetes mode only). Sibling of `workers_per_pod`; controls per-pod fan-out, not the cluster-wide total. Ignored in multiprocessing mode. Sizing guide: the design ratio is 1 record processor per 4 workers.","title":"Recordprocessorsperpod","x-jinja2-supported":true},"workersMin":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Lower bound on worker processes for the adaptive ramp. Distinct from `workers` (the upper bound / target total).","title":"Workersmin","x-jinja2-supported":true},"statsInterval":{"anyOf":[{"maximum":1000.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Interval in seconds between realtime stats publishes (dashboards and the per-tick log block). 0 disables the log block while dashboards continue to poll. Defaults to 5s under --ui dashboard, 30s otherwise. Overrides AIPERF_UI_REALTIME_METRICS_INTERVAL.","title":"Statsinterval","x-jinja2-supported":true}},"title":"RuntimeConfig","type":"object"},"SLAFilter":{"additionalProperties":false,"description":"SLA constraint applied to BO scoring or grid filtering.\n\nA trial is considered feasible iff ``stat(metric_tag) op threshold`` holds.\n``BayesianSearchPlanner`` consumes these for soft-penalty BO scoring and\n``write_search_history`` consumes them for lexicographic\nfeasibility-first best-result selection.\n\nLives in ``aiperf.config`` rather than ``aiperf.search_recipes`` because\n``AdaptiveSearchSweep`` carries a ``list[SLAFilter]`` field; the inverse\nlocation would force a circular import. Re-exported from\n``aiperf.search_recipes._base`` for ergonomic recipe authoring.\n\nExample: enforce p95 TTFT under 200 ms via\n``SLAFilter(metric_tag=\"time_to_first_token\", stat=\"p95\", op=\"lt\", threshold=200.0)``.","properties":{"metricTag":{"description":"Metric tag to filter on, e.g. 'time_to_first_token'. Must match a key in RunResult.summary_metrics produced by the run.","title":"Metrictag","type":"string"},"stat":{"default":"p95","description":"Statistic on the metric to compare against the threshold.","enum":["avg","p1","p5","p10","p25","p50","p75","p90","p95","p99","min","max"],"title":"Stat","type":"string"},"op":{"description":"Comparison operator. Filter passes when stat(metric) op threshold is true.","enum":["lt","le","gt","ge"],"title":"Op","type":"string"},"threshold":{"description":"Numeric threshold the metric statistic is compared against.","title":"Threshold","oneOf":[{"description":"Numeric threshold the metric statistic is compared against.","title":"Threshold","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["metricTag","op","threshold"],"title":"SLAFilter","type":"object"},"SLOTier":{"additionalProperties":false,"description":"Named group of SLA filters representing one service-level objective.\n\nDefines a single tier for multi-tier SLO boundary search. Each tier\ncontains one or more SLA filters that must ALL pass for the tier to be\nconsidered feasible at a given concurrency level.\n\nLives in the config layer (alongside SLAFilter) to avoid circular imports\nbetween config and orchestrator packages. Re-exported from\n``aiperf.orchestrator.search_planner.multi_tier_models`` for ergonomic\norchestrator usage.","properties":{"label":{"description":"Unique tier identifier for output artifacts","title":"Label","type":"string"},"filters":{"description":"SLA filters that must ALL pass for this tier to be feasible","items":{"$ref":"#/$defs/SLAFilter"},"minItems":1,"title":"Filters","type":"array"},"orderingRank":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Rank in detected monotonic ordering (0=strictest); None if unordered","title":"Orderingrank","x-jinja2-supported":true}},"required":["label","filters"],"title":"SLOTier","type":"object"},"SageMakerConfig":{"additionalProperties":false,"description":"SageMaker Runtime routing options.\n\nGrouped rather than flattened onto ``EndpointConfig`` because they only\napply to one transport. Follows the same shape as ``reset_kv_cache`` /\n``server_profiler``.","properties":{"endpointName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Name of the SageMaker endpoint to invoke. Setting this is normally all that is needed: it selects the SageMaker transport, enables SigV4 signing, and derives the runtime base URL from aws_region.","title":"Endpointname"},"targetModel":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker TargetModel value, for multi-model endpoints only: SageMaker rejects requests that carry it on a single-model endpoint. Sent only when set -- it is a relative model-artifact path, so the request's model name is never used in its place. Not sent on streaming requests, which the API does not accept it on.","title":"Targetmodel"},"inferenceComponentName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker InferenceComponentName to target, for endpoints hosting multiple inference components.","title":"Inferencecomponentname"},"targetVariant":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"SageMaker production variant to pin every request to, bypassing the endpoint's traffic split. Useful for benchmarking one variant of an A/B deployment in isolation.","title":"Targetvariant"}},"title":"SageMakerConfig","type":"object"},"SamplingDimension":{"additionalProperties":false,"description":"One dimension of a QMC sampling space.\n\nUsed by `SobolSweep` and `LatinHypercubeSweep`. Distinct from\n`SearchSpaceDimension` (BO) — same path/lo/hi grammar, but native\n`choices` support and explicit `scale` validation.\n\nPath is a dotted path of the form `phases.profiling.concurrency` —\nthe same grammar accepted by `_set_nested_value`.","properties":{"path":{"description":"Dotted-path into BenchmarkConfig.","title":"Path","type":"string"},"lo":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive lower bound (with hi).","title":"Lo","x-jinja2-supported":true},"hi":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Inclusive upper bound (with lo).","title":"Hi","x-jinja2-supported":true},"scale":{"default":"linear","description":"Mapping from unit cube to value.","enum":["linear","log"],"title":"Scale","type":"string"},"kind":{"default":"real","description":"Output type after mapping.","enum":["int","real"],"title":"Kind","type":"string"},"choices":{"anyOf":[{"items":{},"type":"array"},{"type":"null"}],"default":null,"description":"Discrete value list. Mutually exclusive with (lo, hi).","title":"Choices"}},"required":["path"],"title":"SamplingDimension","type":"object"},"ScenarioSweep":{"additionalProperties":false,"description":"Scenario sweep - hand-picked configurations deep-merged with base.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"scenarios","default":"scenarios","description":"Sweep type discriminator.","title":"Type","type":"string"},"runs":{"description":"List of scenario dicts to deep-merge with base config.","items":{"additionalProperties":true,"type":"object"},"minItems":1,"title":"Runs","type":"array"}},"required":["runs"],"title":"ScenarioSweep","type":"object"},"SearchPlannerType":{"enum":["bayesian","monotonic_sla","smooth_isotonic","optuna"],"title":"SearchPlannerType","type":"string"},"SearchSpaceDimension":{"additionalProperties":false,"description":"One dimension of the BO search space.\n\n`path` is a dotted path of the form `phases.profiling.concurrency` —\nthe same grammar accepted by `aiperf.config.sweep._set_nested_value`.","properties":{"path":{"description":"Dotted-path into BenchmarkConfig (e.g. 'phases.profiling.concurrency').","title":"Path","type":"string"},"lo":{"description":"Inclusive lower bound.","title":"Lo","oneOf":[{"description":"Inclusive lower bound.","title":"Lo","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"hi":{"description":"Inclusive upper bound.","title":"Hi","oneOf":[{"description":"Inclusive upper bound.","title":"Hi","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"kind":{"default":"real","description":"Dimension type. 'int' rounds planner suggestions to integers; 'real' keeps floats.","enum":["int","real"],"title":"Kind","type":"string"},"prior":{"default":"uniform","description":"Sampling prior for BO/Optuna planners. 'uniform' draws Sobol/random samples linearly across [lo, hi]; 'log-uniform' draws them log-spaced. Use 'log-uniform' when a parameter spans orders of magnitude (e.g. concurrency in [1, 1000]) so the initial-points phase covers low and high decades evenly. Requires lo > 0. Ignored by planners that do not draw initial points (monotonic, smooth_isotonic) where the search procedure is structurally adaptive rather than sampling-based.","enum":["uniform","log-uniform"],"title":"Prior","type":"string"}},"required":["path","lo","hi"],"title":"SearchSpaceDimension","type":"object"},"SequenceDistributionEntry":{"additionalProperties":false,"description":"Defines a single entry in an ISL/OSL probability distribution.\n\nAIPerf supports multi-modal token length distributions, allowing\ndifferent ISL/OSL combinations with relative frequencies for\nrealistic workload modeling.\n\nYAML Representation:\n sequence_distribution:\n - {isl: 128, osl: 64, probability: 40}\n - {isl: {mean: 512, stddev: 50}, osl: 256, probability: 35}\n - {isl: {mean: 2048, median: 1800}, osl: 512, probability: 25}","examples":[{"isl":128,"osl":64,"probability":40},{"isl":512,"isl_stddev":50,"osl":256,"osl_stddev":25,"probability":35}],"properties":{"isl":{"description":"Input sequence length (tokens). Can be a fixed integer, a {mean, stddev} distribution, or any distribution like {mean: 512, median: 400} for lognormal.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Isl","x-jinja2-supported":true},"islStddev":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Shorthand for ISL standard deviation. If provided when isl is an integer, creates a normal distribution. Cannot be used when isl is already a distribution dict.","title":"Islstddev","x-jinja2-supported":true},"osl":{"description":"Output sequence length (tokens). Can be a fixed integer, a {mean, stddev} distribution, or any typed distribution.","oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"title":"Osl","x-jinja2-supported":true},"oslStddev":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Shorthand for OSL standard deviation. If provided when osl is an integer, creates a normal distribution. Cannot be used when osl is already a distribution dict.","title":"Oslstddev","x-jinja2-supported":true},"probability":{"description":"Relative probability weight for this distribution bucket (0-100). Weights are normalized across all entries. Example: probability=40 means 40%% of requests use this ISL/OSL.","title":"Probability","oneOf":[{"description":"Relative probability weight for this distribution bucket (0-100). Weights are normalized across all entries. Example: probability=40 means 40%% of requests use this ISL/OSL.","maximum":100.0,"minimum":0.0,"title":"Probability","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["isl","osl","probability"],"title":"SequenceDistributionEntry","type":"object"},"ServerMetricsConfig":{"title":"ServerMetricsConfig","description":"Server metrics configuration for Prometheus scraping.\n\nCollects server-side operational metrics (queue depth, KV cache utilization,\nbatch sizes, GPU memory) from Prometheus endpoints exposed by inference servers\nlike vLLM, TensorRT-LLM, or Triton.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9090/metrics\"\n → ServerMetricsConfig(enabled=True, urls=[\"http://localhost:9090/metrics\"])\n - Singular url field: {url: \"...\"}\n → ServerMetricsConfig(urls=[\"...\"]) Can be a URL string or full config object.","anyOf":[{"additionalProperties":false,"description":"Server metrics configuration for Prometheus scraping.\n\nCollects server-side operational metrics (queue depth, KV cache utilization,\nbatch sizes, GPU memory) from Prometheus endpoints exposed by inference servers\nlike vLLM, TensorRT-LLM, or Triton.\n\nAccepts shorthand forms:\n - String URL: \"http://localhost:9090/metrics\"\n → ServerMetricsConfig(enabled=True, urls=[\"http://localhost:9090/metrics\"])\n - Singular url field: {url: \"...\"}\n → ServerMetricsConfig(urls=[\"...\"])","properties":{"enabled":{"default":true,"description":"Enable Prometheus metrics scraping. Set to false to disable.","title":"Enabled","type":"boolean"},"urls":{"description":"Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","items":{"type":"string"},"title":"Urls","type":"array"},"formats":{"description":"Export formats for scraped metrics. Options: json, csv, parquet, jsonl.","items":{"$ref":"#/$defs/ServerMetricsFormat"},"title":"Formats","type":"array"},"discovery":{"$ref":"#/$defs/ServerMetricsDiscoveryConfig","description":"Auto-discovery of Prometheus endpoints in Kubernetes. Discovers pods via Dynamo labels, Prometheus annotations, or custom label selectors."},"url":{"description":"Shorthand for 'urls'. Accepts a single value which will be converted to a list/dict. Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","oneOf":[{"type":"string"},{"description":"Prometheus metrics endpoint URLs to scrape. Typically the /metrics endpoint on inference servers.","items":{"type":"string"},"title":"Urls","type":"array"}],"x-singular-alias-of":"urls"}},"title":"ServerMetricsConfig","type":"object","x-kubernetes-preserve-unknown-fields":true},{"type":"string"}]},"ServerMetricsDiscoveryConfig":{"additionalProperties":false,"description":"Kubernetes-based auto-discovery of inference-server /metrics endpoints.\n\nWhen mode is 'auto' or 'kubernetes', queries the K8s API for pods that\nare recognizable inference servers (vLLM, SGLang, Triton Inference Server,\nTensorRT-LLM, NVIDIA Dynamo). Eligibility (any one is enough):\n1. Dynamo opt-in label: nvidia.com/metrics-enabled=true\n2. AIPerf opt-in annotation: aiperf.nvidia.com/metrics-paths=...\n3. A container image matching a known inference-server signature\n4. User-provided label_selector (server-side filter)\n\nThe broad ``prometheus.io/scrape=true`` annotation is intentionally NOT a\ntrigger: Loki, Grafana, kube-state-metrics, and many platform components\nset it without being inference servers. ``prometheus.io/{port,path,scheme}``\nare still honored to construct the scrape URL when an eligible pod sets them.","properties":{"mode":{"$ref":"#/$defs/ServerMetricsDiscoveryMode","default":"auto","description":"Discovery mode: 'auto' detects environment and tries K8s if in-cluster, 'kubernetes' forces K8s API discovery, 'disabled' uses only explicit URLs."},"labelSelector":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Kubernetes label selector for discovery. Example: 'app=vllm,env=prod'. Applied in addition to built-in Dynamo and Prometheus discovery.","title":"Labelselector"},"namespace":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Kubernetes namespace to search. If not specified, searches the benchmark pod's own namespace (the only namespace the default benchmark RBAC can list pods in). Use '*' to search all namespaces; this requires a cluster-scoped 'pods: list' grant on the benchmark ServiceAccount.","title":"Namespace"},"timeoutSeconds":{"default":30.0,"description":"Maximum seconds to wait for Kubernetes endpoint discovery before proceeding with explicit endpoints only.","title":"Timeoutseconds","oneOf":[{"default":30.0,"description":"Maximum seconds to wait for Kubernetes endpoint discovery before proceeding with explicit endpoints only.","exclusiveMinimum":0,"title":"Timeoutseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"ServerMetricsDiscoveryConfig","type":"object"},"ServerMetricsDiscoveryMode":{"description":"Mode for discovering server metrics endpoints in distributed environments.","enum":["auto","kubernetes","disabled"],"title":"ServerMetricsDiscoveryMode","type":"string"},"ServerMetricsFormat":{"description":"Format options for server metrics export.\n\nControls which output files are generated for server metrics data.\nDefault selection is JSON + CSV (JSONL excluded to avoid large files).","enum":["json","csv","jsonl","parquet"],"title":"ServerMetricsFormat","type":"string"},"ServerProfilerConfig":{"additionalProperties":false,"properties":{"urls":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"description":"Profiler control server URLs. When unset, uses the inference endpoint origins. Paths and queries are ignored.","title":"Urls"},"startBody":{"anyOf":[{"additionalProperties":{"$ref":"#/$defs/JsonValue"},"type":"object"},{"type":"null"}],"default":null,"description":"JSON object sent with profiler start requests. When unset, sends an empty body.","title":"Startbody","x-kubernetes-preserve-unknown-fields":true},"timeoutSeconds":{"anyOf":[{"gt":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Timeout in seconds for profiler start/stop POSTs. When unset, uses 30.0s.","title":"Timeoutseconds","x-jinja2-supported":true},"startPath":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for profiler start. Default: /start_profile.","title":"Startpath"},"stopPath":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Relative path for profiler stop. Default: /stop_profile.","title":"Stoppath"}},"title":"ServerProfilerConfig","type":"object"},"ServiceRunType":{"enum":["kubernetes","multiprocessing"],"title":"ServiceRunType","type":"string"},"SobolSweep":{"additionalProperties":false,"description":"Sobol quasi-Monte-Carlo sweep — low-discrepancy joint coverage.\n\nNote on ``scramble=False``: scipy's raw Sobol sequence starts at the\norigin (the first row is exactly all-zeros), so the first variant\nwill land on ``(lo, lo, ...)`` for every dimension. With small\n``samples`` this produces degenerate corner-only coverage. Prefer\nthe default ``scramble=True`` (Owen-scrambled) unless you have a\nspecific reason to want raw Sobol points.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"samples":{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","title":"Samples","oneOf":[{"description":"Number of variations to draw. Capped at 2**20 (1,048,576) — scipy's Sobol engine fails at 2**30 anyway, and full materialization beyond ~1M variants risks orchestrator OOM.","maximum":1048576,"minimum":2,"title":"Samples","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"seed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"RNG seed for reproducibility. Must be non-negative; scipy's QMC engines reject negative seeds with an opaque low-level error.","title":"Seed","x-jinja2-supported":true},"dimensions":{"description":"Dimensions to sample over.","items":{"$ref":"#/$defs/SamplingDimension"},"minItems":1,"title":"Dimensions","type":"array"},"labelFormat":{"default":"index","description":"Variation-label format. 'index' => '_NNNN' (short, sortable). 'kv' => 'k1=v1, k2=v2' (grid-style).","enum":["index","kv"],"title":"Labelformat","type":"string"},"type":{"const":"sobol","default":"sobol","description":"Sweep type discriminator.","title":"Type","type":"string"},"scramble":{"default":true,"description":"Owen-scrambled Sobol (recommended). False = raw Sobol sequence.","title":"Scramble","type":"boolean"}},"required":["samples","dimensions"],"title":"SobolSweep","type":"object"},"SweepMode":{"description":"Execution order for sweep + confidence composition.\n\nREPEATED: for trial in trials: for value in values\n All values tested per trial before moving to next trial.\n\nINDEPENDENT: for value in values: for trial in trials\n All trials completed per value before moving to next value.","enum":["independent","repeated"],"title":"SweepMode","type":"string"},"SynthesisConfig":{"additionalProperties":false,"description":"Configuration for trace synthesis/transformation.\n\nUsed with mooncake_trace format to transform production trace\ndata before replay. Allows scaling timestamps, token lengths,\nand radix tree structure.","properties":{"speedupRatio":{"default":1.0,"description":"Multiplier for timestamp scaling in synthesized traces. 1.0 = real-time, 2.0 = 2x faster, 0.5 = 2x slower.","title":"Speedupratio","oneOf":[{"default":1.0,"description":"Multiplier for timestamp scaling in synthesized traces. 1.0 = real-time, 2.0 = 2x faster, 0.5 = 2x slower.","exclusiveMinimum":0.0,"title":"Speedupratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"prefixLenMultiplier":{"default":1.0,"description":"Multiplier for core prefix branch lengths in the radix tree. 1.5 means prefix branches are 50%% longer.","title":"Prefixlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for core prefix branch lengths in the radix tree. 1.5 means prefix branches are 50%% longer.","exclusiveMinimum":0.0,"title":"Prefixlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"prefixRootMultiplier":{"default":1,"description":"Number of independent radix trees to distribute traces across. Higher values increase prefix diversity.","title":"Prefixrootmultiplier","oneOf":[{"default":1,"description":"Number of independent radix trees to distribute traces across. Higher values increase prefix diversity.","minimum":1,"title":"Prefixrootmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"promptLenMultiplier":{"default":1.0,"description":"Multiplier for leaf path (unique prompt) lengths. 2.0 means prompts are 2x longer.","title":"Promptlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for leaf path (unique prompt) lengths. 2.0 means prompts are 2x longer.","exclusiveMinimum":0.0,"title":"Promptlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"outputLenMultiplier":{"default":1.0,"description":"Multiplier for output lengths in synthesized traces.","title":"Outputlenmultiplier","oneOf":[{"default":1.0,"description":"Multiplier for output lengths in synthesized traces.","minimum":0.0,"title":"Outputlenmultiplier","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"maxIsl":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum input sequence length filter. Traces with input_length > max_isl are skipped entirely.","title":"Maxisl","x-jinja2-supported":true},"maxOsl":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum output sequence length cap. Traces with output_length > max_osl are capped to this value (not filtered).","title":"Maxosl","x-jinja2-supported":true},"allowDatasetWrap":{"default":false,"description":"Allow the finite weka/agentic trace pool to wrap (reuse distinct roots across concurrency lanes) when concurrency exceeds the number of eligible traces. Defaults to False: over-subscription fails unless this is explicitly enabled or an active cache-bust target already keeps repeated-trace traffic distinct.","title":"Allowdatasetwrap","type":"boolean"}},"title":"SynthesisConfig","type":"object"},"SyntheticDataset":{"additionalProperties":false,"description":"Synthetic dataset configuration.\n\nGenerates prompts programmatically based on token length\nspecifications. Ideal for controlled experiments.","properties":{"endConversationOnTruncation":{"anyOf":[{"type":"boolean"},{"type":"null"}],"default":null,"description":"End a multi-turn conversation as soon as one of its turns is truncated by the output-token cap (the server reports finish_reason=length, stop_reason=max_tokens, or an incomplete Responses API reply), and skip that conversation's remaining turns. The truncated turn itself is still measured. Follow-up prompts assume a complete prior answer, so replaying a cut-off one sends later turns off-distribution. None uses the dataset loader's default: on for SPEED-Bench qualitative datasets (public and custom) and spec_al_mtbench, off otherwise. Rejected at startup for DAG datasets and agentic replay.","title":"Endconversationontruncation"},"systemPrompt":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Verbatim system prompt text, identical across every conversation. Emitted as a system-role message ahead of all turns. When the dataset already carries its own system message, this text is prepended to it rather than replacing it. Tokens are additive: --isl continues to size the generated user prompt only. Mutually exclusive with system_prompt_file and with prefix_prompts.shared_system_length/pool_size/length.","title":"Systemprompt"},"systemPromptFile":{"anyOf":[{"format":"path","type":"string"},{"type":"null"}],"default":null,"description":"Path to a UTF-8 text file holding the verbatim system prompt. Preferred over system_prompt for real production prompts, which are long enough that shell quoting mangles them. Read once at config-validation time so a missing or unreadable file fails at startup. Mutually exclusive with system_prompt.","title":"Systempromptfile"},"name":{"description":"Dataset identifier — used in result file paths. Defaults to 'default' for the singular `dataset:` shorthand.","minLength":1,"title":"Name","type":"string"},"type":{"const":"synthetic","description":"Dataset type discriminator. Must be 'synthetic'.","title":"Type","type":"string"},"entries":{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","title":"Entries","oneOf":[{"default":100,"description":"Total number of unique entries to generate for the dataset. Each entry represents a unique prompt with sampled ISL/OSL. Entries are reused across conversations and turns according to the sampling strategy. Higher values provide more diversity.","minimum":1,"title":"Entries","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"randomSeed":{"anyOf":[{"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Random seed for deterministic dataset generation. When set, makes synthetic prompts, sampling, and other random operations reproducible across runs. Essential for A/B testing and debugging. Overrides global random_seed for this dataset.","title":"Randomseed","x-jinja2-supported":true},"sampling":{"$ref":"#/$defs/DatasetSamplingStrategy","default":"sequential","description":"Strategy for selecting entries from dataset during benchmarking. sequential: iterate in order, wrapping to start after end. random: randomly sample with replacement (entries may repeat). shuffle: random permutation without replacement, re-shuffling after exhaustion."},"prompts":{"anyOf":[{"$ref":"#/$defs/PromptConfig"},{"type":"null"}],"default":null,"description":"Prompt/token length configuration specifying ISL, OSL, sequence distributions, and batch processing settings."},"prefixPrompts":{"anyOf":[{"$ref":"#/$defs/PrefixPromptConfig"},{"type":"null"}],"default":null,"description":"Shared prefix configuration for KV cache testing. Generates prefix prompts that are prepended to user prompts, simulating cached context scenarios."},"turns":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Number of request-response turns per conversation. Can be a fixed integer or {mean, stddev} distribution. Each turn consists of a user message and model response. Set to 1 for single-turn interactions. Multi-turn conversations enable testing of context retention and conversation history handling.","title":"Turns"},"turnDelay":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Delay in milliseconds between consecutive turns within a multi-turn conversation. Can be a fixed value or {mean, stddev} distribution. Simulates user think time between receiving a response and sending the next message. Only applies when turns > 1. Set to 0 for back-to-back turns.","title":"Turndelay"},"turnDelayRatio":{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","title":"Turndelayratio","oneOf":[{"default":1.0,"description":"Multiplier for scaling all turn delays. Applied after mean/stddev calculation: actual_delay = calculated_delay * ratio. Values < 1 speed up conversations, > 1 slow them down. Set to 0 to eliminate delays entirely.","minimum":0.0,"title":"Turndelayratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"images":{"anyOf":[{"$ref":"#/$defs/ImageConfig"},{"type":"null"}],"default":null,"description":"Synthetic image configuration for multimodal vision-language testing."},"audio":{"anyOf":[{"$ref":"#/$defs/AudioConfig"},{"type":"null"}],"default":null,"description":"Synthetic audio configuration for multimodal speech/audio testing."},"video":{"anyOf":[{"$ref":"#/$defs/VideoConfig"},{"type":"null"}],"default":null,"description":"Synthetic video configuration for multimodal video understanding testing."},"rankings":{"anyOf":[{"$ref":"#/$defs/RankingsConfig"},{"type":"null"}],"default":null,"description":"Rankings/reranking configuration for generating query-passage pairs. Only relevant for rankings endpoint types."},"isl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.isl; hoisted into prompts before validation.","title":"Isl"},"osl":{"anyOf":[{"oneOf":[{"$ref":"#/$defs/FixedDistribution"},{"$ref":"#/$defs/NormalDistribution"},{"$ref":"#/$defs/LogNormalDistribution"},{"$ref":"#/$defs/MultimodalDistribution"},{"$ref":"#/$defs/EmpiricalDistribution"}]},{"type":"null"}],"default":null,"description":"Shorthand for prompts.osl; hoisted into prompts before validation.","title":"Osl"}},"required":["name","type"],"title":"SyntheticDataset","type":"object"},"TcpCommunicationConfig":{"additionalProperties":false,"description":"TCP socket communication configuration.\n\nFor distributed deployments across machines.\nProvides detailed port configuration for all ZMQ proxies.","properties":{"type":{"const":"tcp","description":"Communication type. Must be 'tcp'.","title":"Type","type":"string"},"host":{"default":"127.0.0.1","description":"Host address for TCP communication. Use 0.0.0.0 to listen on all interfaces.","title":"Host","type":"string"},"recordsPort":{"default":5557,"description":"Port for records push/pull communication.","title":"Recordsport","oneOf":[{"default":5557,"description":"Port for records push/pull communication.","maximum":65535,"minimum":1,"title":"Recordsport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditRouterPort":{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","title":"Creditrouterport","oneOf":[{"default":5564,"description":"Port for credit router (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Creditrouterport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"controlPort":{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","title":"Controlport","oneOf":[{"default":5667,"description":"Port for control channel (ROUTER-DEALER).","maximum":65535,"minimum":1,"title":"Controlport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"creditReturnPushPullPort":{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","title":"Creditreturnpushpullport","oneOf":[{"default":5669,"description":"Port for the credit-return PUSH/PULL fan-in channel.","maximum":65535,"minimum":1,"title":"Creditreturnpushpullport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"eventBusProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Event bus proxy ports (XPUB/XSUB)."},"datasetManagerProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Dataset manager proxy ports (DEALER/ROUTER)."},"rawInferenceProxy":{"$ref":"#/$defs/TcpProxyConfig","description":"Raw inference proxy ports (PUSH/PULL)."}},"required":["type"],"title":"TcpCommunicationConfig","type":"object"},"TcpProxyConfig":{"additionalProperties":false,"description":"TCP proxy port configuration for a single ZMQ proxy.","properties":{"frontendPort":{"description":"TCP port for proxy frontend.","title":"Frontendport","oneOf":[{"description":"TCP port for proxy frontend.","maximum":65535,"minimum":1,"title":"Frontendport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"backendPort":{"description":"TCP port for proxy backend.","title":"Backendport","oneOf":[{"description":"TCP port for proxy backend.","maximum":65535,"minimum":1,"title":"Backendport","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"required":["frontendPort","backendPort"],"title":"TcpProxyConfig","type":"object"},"TemplateConfig":{"additionalProperties":false,"description":"Configuration for custom template-based endpoints.\n\nWhen endpoint type is \"template\", this configures how requests\nare formatted and responses are parsed.","properties":{"body":{"description":"Jinja2 template string for request body. Variables: {{prompt}}, {{max_tokens}}, {{model}}, {{messages}}, etc.","title":"Body","type":"string"},"responseField":{"default":"text","description":"JSON path to extract response text from API response. Use dot notation for nested fields: 'choices.0.message.content'.","title":"Responsefield","type":"string"}},"required":["body"],"title":"TemplateConfig","type":"object"},"TimingMode":{"enum":["fixed_schedule","request_rate","adaptive_scale","user_centric_rate","agentic_replay"],"title":"TimingMode","type":"string"},"TokenizerConfig":{"additionalProperties":false,"description":"Tokenizer configuration for token counting and prompt generation.\n\nAIPerf uses a HuggingFace tokenizer for accurate token counting,\nwhich is essential for ISL/OSL enforcement and metrics calculation.","properties":{"name":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"HuggingFace tokenizer identifier, local filesystem path, or `builtin` for a zero-network-access tokenizer backed by tiktoken (o200k_base encoding). Should match the model's tokenizer for accurate token counts. If `--tokenizer` is not set and the model name looks like an obvious placeholder (e.g. `mock-model`, `test-model`, `fake-model`), AIPerf substitutes `builtin` automatically and emits a warning. Example: 'meta-llama/Llama-3.1-8B-Instruct'","title":"Name"},"revision":{"default":"main","description":"Model revision to use: branch name, tag, or commit hash. Use for version pinning to ensure reproducibility.","title":"Revision","type":"string"},"trustRemoteCode":{"default":false,"description":"Allow execution of custom tokenizer code from the repository. Required for some models but poses security risk. Only enable for trusted sources.","title":"Trustremotecode","type":"boolean"},"applyChatTemplate":{"default":false,"description":"Apply the HuggingFace tokenizer's chat template when counting input tokens. When enabled: synthetic ISL is compensated for chat-template wrapping (BOS, role headers, EOT, generation-prompt suffix) and the record processor reports ISL using `apply_chat_template(tokenize=True, add_generation_prompt=True)` for chat-shape payloads. When disabled (default), both paths use bare-text encoding, so reported ISL matches the prompt content the user asked for and ignores template overhead. Requires an HF tokenizer with a chat template configured; no-ops on tiktoken / un-templated models.","title":"Applychattemplate","type":"boolean"}},"title":"TokenizerConfig","type":"object"},"TokenizerOverride":{"additionalProperties":false,"description":"Per-model tokenizer override configuration.\n\nAllows specifying a different tokenizer for a specific model,\nuseful when models require specialized tokenization.","properties":{"name":{"description":"HuggingFace tokenizer identifier or local filesystem path.","title":"Name","type":"string"}},"required":["name"],"title":"TokenizerOverride","type":"object"},"TransportType":{"enum":["http","sagemaker"],"title":"TransportType","type":"string"},"UIType":{"enum":["dashboard","none","simple"],"title":"UIType","type":"string"},"URLSelectionStrategy":{"enum":["round_robin"],"title":"URLSelectionStrategy","type":"string"},"UserCentricPhase":{"additionalProperties":false,"description":"N concurrent users sharing a global request rate.\n\nRequires multi-turn conversations. Each user gets a proportional\nshare of the global ``rate``.","properties":{"adaptiveScale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]},"adaptiveSustainDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds to sustain load near the discovered adaptive scale boundary.","title":"Adaptivesustainduration","x-jinja2-supported":true},"adaptiveAssessmentPeriod":{"anyOf":[{"minimum":1.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Duration in seconds for each adaptive scale SLA assessment window.","title":"Adaptiveassessmentperiod","x-jinja2-supported":true},"adaptiveMinCompletedRequests":{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","title":"Adaptivemincompletedrequests","oneOf":[{"default":1,"description":"Minimum completed requests needed before an adaptive SLA window can make a decision.","minimum":1,"title":"Adaptivemincompletedrequests","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlVariable":{"default":"concurrency","description":"Named adaptive control variable.","enum":["concurrency","prefill_concurrency","request_rate","users"],"title":"Adaptivecontrolvariable","type":"string"},"adaptiveControlMin":{"default":1,"description":"Minimum adaptive control value.","title":"Adaptivecontrolmin","oneOf":[{"default":1,"description":"Minimum adaptive control value.","exclusiveMinimum":0,"title":"Adaptivecontrolmin","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveControlMax":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Maximum adaptive control value. Inferred from the phase target when omitted.","title":"Adaptivecontrolmax","x-jinja2-supported":true},"adaptiveScaleStrategyType":{"const":"ramp_until_fail","default":"ramp_until_fail","description":"Adaptive scale controller strategy. v1 supports ramp_until_fail.","title":"Adaptivescalestrategytype","type":"string"},"adaptiveScaleStepPolicy":{"default":"sla_margin","description":"Adaptive scale increase policy. sla_margin uses normalized SLA margin to choose larger steps when far from the boundary; fixed_percent_step uses a fixed percentage of the current control value.","enum":["sla_margin","fixed_percent_step"],"title":"Adaptivescalesteppolicy","type":"string"},"adaptiveScaleBaseStep":{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","title":"Adaptivescalebasestep","oneOf":[{"default":10,"description":"Minimum adaptive scale step for SLA-margin policy.","minimum":1,"title":"Adaptivescalebasestep","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleMaxStepMultiplier":{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","title":"Adaptivescalemaxstepmultiplier","oneOf":[{"default":4,"description":"Maximum base-step multiplier for SLA-margin policy.","minimum":1,"title":"Adaptivescalemaxstepmultiplier","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptiveScaleStepPercent":{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","title":"Adaptivescalesteppercent","oneOf":[{"default":25.0,"description":"Percent of current concurrency used by fixed-percent adaptive scaling.","exclusiveMinimum":0,"title":"Adaptivescalesteppercent","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"name":{"description":"Unique workflow label for this phase, such as 'baseline_traffic', 'cancellation_stress', or 'recovery_traffic'. This is distinct from phase kind: multiple phases may share kind='profiling' while each keeps a different name. Used in logs, status, sweep targeting, artifact paths, and result file naming. Must be a strict identifier: letters, numbers, underscores, and hyphens; must start with a letter or underscore.","pattern":"^[A-Za-z_][A-Za-z0-9_-]*$","title":"Name","type":"string"},"kind":{"anyOf":[{"enum":["warmup","profiling"],"type":"string"},{"type":"null"}],"default":null,"description":"Semantic runtime role for the phase. Only 'warmup' and 'profiling' are valid kinds because the credit/results pipeline distinguishes those two roles. This field is nullable only as an input compatibility bridge: legacy canonical names ('warmup' and 'profiling') infer kind during normalization, while validated phases always carry a concrete kind.","title":"Kind"},"type":{"const":"user_centric","description":"N users sharing a global request rate.","title":"Type","type":"string"},"excludeFromResults":{"default":false,"description":"Exclude this phase's metrics from final results. Forced by phase kind: kind='warmup' is always excluded, kind='profiling' is always included. Explicitly setting this field to a value inconsistent with the phase kind is rejected.","title":"Excludefromresults","type":"boolean"},"requests":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many requests sent (must be >= 1).","title":"Requests","x-jinja2-supported":true},"duration":{"anyOf":[{"type":"string","pattern":"^(?:\\d+(?:\\.\\d+)?\\s*(?:[sS]|[sS][eE][cC]|[mM]|[mM][iI][nN]|[hH]|[hH][rR]|[hH][oO][uU][rR])?|[iI][nN][fF](?:[iI][nN][iI][tT][yY])?)$","description":"Duration string (e.g., '30s', '5M', '2h', 'inf')."},{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this time elapsed (must be > 0). Supports: 300, '5m', '2h'. Accepts number (seconds) or string like '30s', '5m', '2h'.","gt":0,"title":"Duration","x-jinja2-supported":true},"sessions":{"anyOf":[{"minimum":1,"type":"integer"},{"const":"all","type":"string"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Stop after this many sessions completed (must be >= 1). 'all' runs one full pass: the session count resolves at run time to the number of conversations the dataset loaded.","title":"Sessions","x-jinja2-supported":true},"concurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent in-flight requests (must be >= 1). For concurrency type: primary control. For rate types: acts as a cap.","title":"Concurrency","x-jinja2-supported":true},"concurrencyRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"prefillConcurrency":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Max concurrent requests in prefill stage (must be >= 1). Limits requests before first token received.","title":"Prefillconcurrency","x-jinja2-supported":true},"prefillRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp prefill_concurrency from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"gracePeriod":{"anyOf":[{"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Seconds to wait for in-flight requests after duration expires (must be >= 0). Requires 'duration' to be set. Supports: 30, '30s', '2m'.","ge":0,"title":"Graceperiod","x-jinja2-supported":true},"cancellation":{"anyOf":[{"$ref":"#/$defs/CancellationConfig"},{"type":"null"}],"default":null,"description":"Request cancellation testing configuration."},"sla":{"description":"SLA filters evaluated by adaptive load controllers. Accepts the canonical list form or compact metric/stat/op mapping.","title":"Sla","anyOf":[{"description":"SLA filters evaluated by adaptive load controllers.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Sla","type":"array"},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"seamless":{"default":false,"description":"Start this phase immediately when previous phase stops, without waiting for in-flight requests to complete. Cannot be True for the first phase.","title":"Seamless","type":"boolean"},"timingMode":{"anyOf":[{"$ref":"#/$defs/TimingMode"},{"type":"null"}],"default":null,"description":"Explicit timing-strategy override for this phase. When set, it WINS over the timing mode derived from ``type`` in ``aiperf.timing.config._phase_timing_mode`` (read via ``_is_agentic_replay`` and ``_build_profiling_config``). This is how a benchmark-scenario invariant-lock stamps AGENTIC_REPLAY onto the profiling phase(s): the scenario validator sets ``phase.timing_mode = TimingMode.AGENTIC_REPLAY`` and the credit pipeline then routes the phase through the agentic-replay strategy. Leave None for normal/dag_jsonl runs so the phase-type mapping applies (REQUEST_RATE / FIXED_SCHEDULE / USER_CENTRIC_RATE). Distinct from the sweep ``scenarios`` strategy."},"failedRequestThreshold":{"anyOf":[{"maximum":1.0,"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Abort the run early when (failed_records / total_records) exceeds this ratio. Default None disables the check. Only PROFILING-phase records count toward the ratio. A grace floor of max(concurrency, 10) records must accumulate before the check is armed, so a single early failure cannot kill the run. When the threshold is exceeded a PROFILE_CANCEL is broadcast: in-flight requests drain via the normal cancel path, partial results are still aggregated, and the run exits non-zero. Pairs with the AGENTIC_REPLAY context-overflow drop in record_processor_service so the rate measures real failures only.","title":"Failedrequestthreshold","x-jinja2-supported":true},"trajectoryStartMinRatio":{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","title":"Trajectorystartminratio","oneOf":[{"default":0.25,"description":"AGENTIC_REPLAY only: lower bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). Sampled per trajectory at trajectory-build time; deterministic given --random-seed.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartminratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"trajectoryStartMaxRatio":{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","title":"Trajectorystartmaxratio","oneOf":[{"default":0.75,"description":"AGENTIC_REPLAY only: upper bound (inclusive) on the random start position within each trajectory, expressed as a fraction of the trace's recorded wall-clock duration (timestamped traces) or its total turn count (legacy timestamp-less traces). For the timestamp-less path the effective per-trace ceiling is min(int(max_ratio * n), n - 2) so at least one profile turn remains after warmup.","maximum":1.0,"minimum":0.0,"title":"Trajectorystartmaxratio","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"burstPhaseStarts":{"default":false,"description":"AGENTIC_REPLAY only: collapse the WARMUP-start and PROFILING-start dispatches into synchronized bursts instead of preserving recorded spacing. By default (False), WARMUP requests are aligned globally so every trajectory reaches t* together. PROFILING subtracts one phase-wide minimum from every first request offset: the earliest request starts immediately and all other trajectories retain their recorded spacing. Pass --burst-phase-starts to subtract a separate minimum per lane, making every lane start immediately. Subsequent inter-turn delays are timing-faithful in either mode.","title":"Burstphasestarts","type":"boolean"},"systemIdleGapCapSeconds":{"anyOf":[{"minimum":0.0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: maximum time in seconds the replay may remain globally idle while future requests are scheduled. When no requests are in flight or ready, all pending request timers shift earlier by the same amount so the next request arrives within this limit. Per-trace timing, timer order, and relative spacing are otherwise preserved. None disables the global idle cap.","title":"Systemidlegapcapseconds","x-jinja2-supported":true},"agenticCacheWarmupDuration":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: additional cache-pressure warmup duration in seconds. After the normal snapshot warmup drains, AIPerf continues the live trajectories without recorded idle delays and with one-token outputs for this long, then drains and resumes profiling from the resulting trajectory state. Read off the profiling phase by ``timing.config._build_agentic_warmup_config``. Mutually exclusive with warmup_requests_per_lane. None disables it.","title":"Agenticcachewarmupduration","x-jinja2-supported":true},"warmupRequestsPerLane":{"anyOf":[{"exclusiveMinimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: deterministic cache-pressure warmup request budget per concurrency lane, additional to mandatory snapshot primers. The total warmup request target is the primer count plus this value multiplied by the number of live trajectory lanes. Mutually exclusive with agentic_cache_warmup_duration.","title":"Warmuprequestsperlane","x-jinja2-supported":true},"agenticWarmupGracePeriod":{"anyOf":[{"minimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"AGENTIC_REPLAY only: grace period in seconds the auto-synthesized warmup barrier waits for in-flight priming requests after the warmup burst sends. Read off the profiling phase by ``timing.config._build_agentic_warmup_config`` (the agentic warmup is not a user-declared phase, so it does not inherit ``--warmup-grace-period``, which requires ``--warmup-duration``). None waits indefinitely (the agentic warmup must complete every primed trajectory before profiling starts).","title":"Agenticwarmupgraceperiod","x-jinja2-supported":true},"rate":{"anyOf":[{"exclusiveMinimum":0,"type":"number"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Target request rate in requests per second. Required unless rate_series is set.","title":"Rate","x-jinja2-supported":true},"rateRamp":{"anyOf":[{"$ref":"#/$defs/RampConfig"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Ramp rate from lower value. Can be number (seconds) or {duration, strategy}.","x-jinja2-supported":true},"users":{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","title":"Users","oneOf":[{"description":"Number of simulated concurrent users (must be >= 1). Requests distributed across users to achieve global rate.","minimum":1,"title":"Users","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"adaptive_scale":{"description":"Enable single-run adaptive scale control for this phase. Accepts either a boolean or a nested adaptive-scale configuration block.","title":"Adaptivescale","anyOf":[{"default":false,"description":"Enable single-run adaptive scale control for this phase.","title":"Adaptivescale","type":"boolean"},{"$ref":"#/$defs/AdaptiveScaleNestedBlock"}]}},"required":["name","type","users","rate"],"title":"UserCentricPhase","type":"object","allOf":[{"required":["rate"],"properties":{"rate":{"not":{"type":"null"}}}}]},"UserFile":{"description":"One user-declared output file rendered into the run directory before benchmark start.\n\nPath is relative to the run directory; subdirectories are allowed; absolute\npaths and any segment equal to '..' are rejected. Content is rendered with\njinja2 against a documented context (variables: + system-injected names).","properties":{"path":{"description":"Output path relative to the run directory. Subdirectories allowed. Absolute paths and any segment equal to '..' are rejected.","title":"Path","type":"string"},"format":{"anyOf":[{"enum":["json","yaml","text"],"type":"string"},{"type":"null"}],"default":null,"description":"Serialization format. If omitted: 'text' when content is a string, 'json' otherwise.","title":"Format"},"content":{"description":"Templated value. Structured (dict/list/scalar) for json/yaml; string for text. Jinja2 expressions in any string leaf are rendered with the user_files context.","title":"Content","x-kubernetes-preserve-unknown-fields":true}},"required":["path","content"],"title":"UserFile","type":"object"},"VideoAudioCodec":{"description":"Audio codecs for embedding audio in synthetic video files.","enum":["aac","libvorbis","libopus"],"title":"VideoAudioCodec","type":"string"},"VideoAudioConfig":{"additionalProperties":false,"description":"Configuration for embedding an audio track in synthetic video files.","properties":{"sampleRate":{"default":44.1,"description":"Audio sample rate in kHz for the embedded audio track. Common values: 8 (telephony), 16 (speech), 44.1 (CD quality), 48 (professional). Higher sample rates increase audio fidelity and file size.","title":"Samplerate","oneOf":[{"default":44.1,"description":"Audio sample rate in kHz for the embedded audio track. Common values: 8 (telephony), 16 (speech), 44.1 (CD quality), 48 (professional). Higher sample rates increase audio fidelity and file size.","maximum":96.0,"minimum":8.0,"title":"Samplerate","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"channels":{"default":0,"description":"Number of audio channels to embed in generated video files. 0 = disabled (no audio track, default), 1 = mono, 2 = stereo. When set to 1 or 2, a Gaussian noise audio track matching the video duration is muxed into each video via FFmpeg.","title":"Channels","oneOf":[{"default":0,"description":"Number of audio channels to embed in generated video files. 0 = disabled (no audio track, default), 1 = mono, 2 = stereo. When set to 1 or 2, a Gaussian noise audio track matching the video duration is muxed into each video via FFmpeg.","maximum":2,"minimum":0,"title":"Channels","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"codec":{"anyOf":[{"$ref":"#/$defs/VideoAudioCodec"},{"type":"null"}],"default":null,"description":"Audio codec for the embedded audio track. If not specified, auto-selects based on video format: libopus for MP4, libvorbis for WebM. Options: libvorbis, libopus, aac. The AIPerf container ships only libvorbis and libopus; aac requires an FFmpeg build that includes an AAC encoder. libopus always encodes at 48 kHz, so any --video-audio-sample-rate is resampled during muxing."},"depth":{"description":"Audio bit depth for the embedded audio track. Supported values: 8, 16, 24, or 32 bits. Higher bit depths provide greater dynamic range but increase file size.","default":16,"oneOf":[{"default":16,"description":"Audio bit depth for the embedded audio track. Supported values: 8, 16, 24, or 32 bits. Higher bit depths provide greater dynamic range but increase file size.","enum":[8,16,24,32],"title":"Depth","type":"integer"},{"type":"string","enum":["8","16","24","32"]},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true}},"title":"VideoAudioConfig","type":"object"},"VideoConfig":{"additionalProperties":false,"description":"Configuration for synthetic video generation in multimodal datasets.\n\nControls the generation of synthetic videos for video understanding\nmodel benchmarking. Requires FFmpeg for video generation.","properties":{"batchSize":{"default":0,"description":"Number of video files to include in each multimodal request. Supported with chat endpoint type for video understanding models. Set to 0 to disable video inputs. Higher batch sizes significantly increase request payload size.","title":"Batchsize","oneOf":[{"default":0,"description":"Number of video files to include in each multimodal request. Supported with chat endpoint type for video understanding models. Set to 0 to disable video inputs. Higher batch sizes significantly increase request payload size.","minimum":0,"title":"Batchsize","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"duration":{"default":1.0,"description":"Duration in seconds for each generated video clip. Combined with fps, determines total frame count (frames = duration * fps). Longer durations increase file size and processing time. Typical values: 1-10 seconds for testing.","title":"Duration","oneOf":[{"default":1.0,"description":"Duration in seconds for each generated video clip. Combined with fps, determines total frame count (frames = duration * fps). Longer durations increase file size and processing time. Typical values: 1-10 seconds for testing.","exclusiveMinimum":0.0,"title":"Duration","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"fps":{"default":4,"description":"Frames per second for generated video. Higher FPS creates smoother video but increases frame count and file size. Common values: 4 (minimal, recommended for Cosmos models), 24 (cinematic), 30 (standard), 60 (high frame rate). Total frames = duration * fps.","title":"Fps","oneOf":[{"default":4,"description":"Frames per second for generated video. Higher FPS creates smoother video but increases frame count and file size. Common values: 4 (minimal, recommended for Cosmos models), 24 (cinematic), 30 (standard), 60 (high frame rate). Total frames = duration * fps.","minimum":1,"title":"Fps","type":"integer"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"width":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Video frame width in pixels. Determines video resolution and file size. Common values: 640 (SD), 1280 (HD), 1920 (Full HD). If not specified, uses codec/format defaults.","title":"Width","x-jinja2-supported":true},"height":{"anyOf":[{"minimum":1,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Video frame height in pixels. Combined with width determines aspect ratio and total pixel count per frame. Common values: 480 (SD), 720 (HD), 1080 (Full HD). If not specified, uses codec/format defaults.","title":"Height","x-jinja2-supported":true},"format":{"$ref":"#/$defs/VideoFormat","default":"webm","description":"Container format for generated video files. webm: VP9 codec, BSD-licensed, recommended for open-source workflows. mp4: widely compatible container, VP9 by default. Format affects compatibility, file size, and encoding options."},"codec":{"default":"libvpx-vp9","description":"Video codec for encoding. Common options: libvpx-vp9 (CPU, BSD-licensed, default), libvpx (CPU, BSD-licensed, VP8). Any codec the local FFmpeg supports can be used, but the AIPerf container ships a minimal FFmpeg build limited to VP8/VP9 video with Vorbis/Opus audio; codecs such as libx264 or h264_nvenc require an FFmpeg build that includes them.","title":"Codec","type":"string"},"synthType":{"$ref":"#/$defs/VideoSynthType","default":"moving_shapes","description":"Algorithm for generating synthetic video content. Different types produce different visual patterns for testing. Content doesn't affect semantic meaning but may impact encoding efficiency and file size."},"audio":{"$ref":"#/$defs/VideoAudioConfig","description":"Audio track configuration for embedding audio in generated videos."}},"title":"VideoConfig","type":"object"},"VideoFormat":{"description":"Video container formats for synthetic video generation.","enum":["mp4","webm"],"title":"VideoFormat","type":"string"},"VideoSynthType":{"enum":["moving_shapes","grid_clock","noise"],"title":"VideoSynthType","type":"string"},"WandbConfig":{"additionalProperties":false,"dependentRequired":{"entity":["project"],"runName":["project"],"tags":["project"]},"description":"Weights & Biases run-upload configuration.","properties":{"project":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases project name. Setting this enables wandb export.","title":"Project"},"entity":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases entity (team or user). Defaults to the API key's default entity.","title":"Entity"},"runName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Weights & Biases run name.","title":"Runname"},"tags":{"anyOf":[{"items":{"type":"string"},"type":"array"},{"type":"null"}],"default":null,"description":"Additional Weights & Biases run tags to attach on upload.","title":"Tags"}},"title":"WandbConfig","type":"object"},"ZipSweep":{"additionalProperties":false,"description":"Zip sweep — parameters paired element-wise (lockstep).\n\nLike ``GridSweep`` but uses ``zip(strict=True)`` instead of Cartesian\nproduct. All parameter lists must have identical length. Use when you\nwant N runs each setting a coordinated tuple of fields, without the\nNxM blow-up of grid. Canonical use case: paired ISL/OSL.","properties":{"cooldownSeconds":{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","title":"Cooldownseconds","oneOf":[{"default":0.0,"description":"Cooldown between sweep variations (or BO iterations).","minimum":0,"title":"Cooldownseconds","type":"number"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"x-jinja2-supported":true},"slaFilters":{"description":"SLA constraints applied at trial scoring / aggregation time.","items":{"$ref":"#/$defs/SLAFilter"},"title":"Slafilters","type":"array"},"postProcess":{"anyOf":[{"$ref":"#/$defs/PostProcessSpec"},{"type":"null"}],"default":null,"description":"Optional post-aggregation handler emitted by a search recipe."},"recipeName":{"anyOf":[{"type":"string"},{"type":"null"}],"default":null,"description":"Source recipe name (audit trail when expanded via --search-recipe). Used by the live Pareto plot path to look up the recipe's pareto_axes after expansion has stripped the originating --search-recipe flag.","title":"Recipename"},"iterationOrder":{"$ref":"#/$defs/SweepMode","default":"repeated","description":"Trial/variation iteration order. 'repeated': trials outer, variations inner. 'independent': variations outer, trials inner. Inapplicable to adaptive_search."},"sameSeed":{"default":false,"description":"If True, every variation reuses the envelope's random_seed. If False, each variation derives `random_seed + variation.index`.","title":"Sameseed","type":"boolean"},"type":{"const":"zip","default":"zip","description":"Sweep type discriminator.","title":"Type","type":"string"},"parameters":{"additionalProperties":{"items":{},"type":"array"},"description":"Parameters to sweep in lockstep: dot-notation path -> list of values. All lists must have equal length.","minProperties":1,"title":"Parameters","type":"object"}},"required":["parameters"],"title":"ZipSweep","type":"object"},"AdaptiveScaleNestedBlock":{"type":"object","additionalProperties":false,"properties":{"enabled":{"anyOf":[{"type":"boolean"},{"type":"string","enum":["true","false","yes","no","on","off","1","0"]}]},"control":{"type":"object","additionalProperties":false,"properties":{"variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"strategy":{"type":"object","additionalProperties":false,"properties":{"type":{"type":"string","enum":["ramp_until_fail"]},"stepPolicy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"step_policy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"baseStep":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"base_step":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxStepMultiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_step_multiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"stepPercent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"step_percent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"sla":{"anyOf":[{"type":"array","items":{"$ref":"#/$defs/SLAFilter"}},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"controlVariable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"control_variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"controlMin":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"controlMax":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"window":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessmentPeriod":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessment_period":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minCompletedRequests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_completed_requests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustainDuration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustain_duration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"DisabledAdaptiveScaleNestedBlock":{"type":"object","additionalProperties":false,"properties":{"enabled":{"anyOf":[{"const":false},{"type":"string","enum":["false","no","off","0"]}]},"control":{"type":"object","additionalProperties":false,"properties":{"variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"strategy":{"type":"object","additionalProperties":false,"properties":{"type":{"type":"string","enum":["ramp_until_fail"]},"stepPolicy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"step_policy":{"type":"string","enum":["sla_margin","fixed_percent_step"]},"baseStep":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"base_step":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxStepMultiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_step_multiplier":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"stepPercent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"step_percent":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}}},"sla":{"anyOf":[{"type":"array","items":{"$ref":"#/$defs/SLAFilter"}},{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":"object","additionalProperties":{"type":["number","string"]}}},"description":"Compact SLA mapping: metric -> stat -> operator -> threshold."}]},"controlVariable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"control_variable":{"type":"string","enum":["concurrency","prefill_concurrency","request_rate","users"]},"controlMin":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_min":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"controlMax":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"control_max":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"maxConcurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"max_concurrency":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"window":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessmentPeriod":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"assessment_period":{"anyOf":[{"type":"integer","minimum":1},{"type":"number","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"minCompletedRequests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"min_completed_requests":{"oneOf":[{"type":"integer","minimum":1},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustainDuration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]},"sustain_duration":{"anyOf":[{"type":"integer","exclusiveMinimum":0},{"type":"number","exclusiveMinimum":0},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}]}},"required":["enabled"]}},"additionalProperties":false,"description":"JSON Schema for AIPerf YAML configuration files. This schema enables IDE autocompletion and validation for AIPerf benchmark configurations. Schema version: 2.0.0","properties":{"schemaVersion":{"const":"2.0","default":"2.0","description":"AIPerf config schema version. Pinned to '2.0' for the envelope/body shape (envelope keys at top level, swept body under ``benchmark``).","title":"Schemaversion","type":"string"},"benchmark":{"$ref":"#/$defs/BenchmarkConfig","description":"Benchmark workload (the swept body)."},"sweep":{"anyOf":[{"discriminator":{"propertyName":"type","mapping":{"grid":"#/$defs/GridSweep","zip":"#/$defs/ZipSweep","scenarios":"#/$defs/ScenarioSweep","adaptive_search":"#/$defs/AdaptiveSearchSweep","sobol":"#/$defs/SobolSweep","latin_hypercube":"#/$defs/LatinHypercubeSweep"}},"anyOf":[{"$ref":"#/$defs/GridSweep"},{"$ref":"#/$defs/ZipSweep"},{"$ref":"#/$defs/ScenarioSweep"},{"$ref":"#/$defs/AdaptiveSearchSweep"},{"$ref":"#/$defs/SobolSweep"},{"$ref":"#/$defs/LatinHypercubeSweep"}]},{"type":"null"}],"default":null,"description":"Sweep configuration for parameter exploration. Supports grid (Cartesian product), zip (lock-stepped axes), scenarios (hand-picked), sobol / latin_hypercube (QMC sampling), and adaptive_search (Bayesian / monotonic outer-loop) strategies.","title":"Sweep"},"multiRun":{"$ref":"#/$defs/MultiRunConfig","description":"Multi-run benchmarking configuration for statistical reporting. When num_runs > 1, executes multiple runs and computes aggregate statistics."},"plot":{"default":null,"description":"Plot configuration for this run/sweep. When set, ``~/.aiperf/plot_config.yaml`` is ignored. May be a bare-string path to a plot YAML (resolved relative to the AIPerf config file's directory, or absolute), or an inline mapping mirroring ``default_plot_config.yaml``. Presence implies ``artifacts.auto_plot=True`` unless explicitly set False.","title":"Plot"},"variables":{"additionalProperties":true,"description":"User-defined values exposed to Jinja2 in `{{ ... }}` expressions during config load. Cross-variation: scenario `runs[i].variables:` deep-merge over this base. Preserved on the resolved config so run-time renderers can resolve them again.","title":"Variables","type":"object"},"randomSeed":{"anyOf":[{"minimum":0,"type":"integer"},{"type":"null"},{"type":"string","pattern":".*\\{\\{.*\\}\\}.*","description":"Jinja2 template (e.g., '{{ variable }}')."},{"type":"string","pattern":".*\\$\\{[A-Za-z_][A-Za-z0-9_]*(?::[^}]*)?\\}.*","description":"Environment variable (e.g., '${VAR}' or '${VAR:default}')."}],"default":null,"description":"Global random seed for reproducibility. Base seed for per-variation derivation in sweep mode (variation N gets base + N). Must be non-negative; numpy and scipy reject negative seeds.","title":"Randomseed","x-jinja2-supported":true},"noSweepTable":{"default":false,"description":"Suppress the per-cell streaming sweep table during multi-variation sweeps. Auto-suppressed when stdout is non-interactive, when the dashboard UI is active, or for single-cell sweeps.","title":"Nosweeptable","type":"boolean"}},"required":["benchmark"],"title":"AIPerf Configuration Schema","type":"object","$comment":"Auto-generated by tools/generate_config_schema.py. Do not edit manually.","x-taplo":{"initKeys":["benchmark","sweep","multiRun","variables","randomSeed"]}} diff --git a/src/aiperf/endpoints/anthropic_messages.py b/src/aiperf/endpoints/anthropic_messages.py index 1031eabf61..ffe541dddd 100644 --- a/src/aiperf/endpoints/anthropic_messages.py +++ b/src/aiperf/endpoints/anthropic_messages.py @@ -446,8 +446,7 @@ def get_endpoint_headers(self, request_info: RequestInfo) -> dict[str, str]: """Get Anthropic-specific headers using x-api-key auth.""" cfg = self.model_endpoint.endpoint headers: dict[str, str] = {"content-type": "application/json"} - if cfg.headers: - headers.update(cfg.headers) + headers.update(cfg.get_headers(request_info.url_index)) if cfg.api_key and not cfg.auth_type: headers["x-api-key"] = cfg.api_key headers.setdefault("anthropic-version", _ANTHROPIC_VERSION) diff --git a/src/aiperf/endpoints/base_endpoint.py b/src/aiperf/endpoints/base_endpoint.py index 92627bd445..632d0477f5 100644 --- a/src/aiperf/endpoints/base_endpoint.py +++ b/src/aiperf/endpoints/base_endpoint.py @@ -44,7 +44,7 @@ def __init__(self, model_endpoint: ModelEndpointInfo, **kwargs): def get_endpoint_headers(self, request_info: RequestInfo) -> dict[str, str]: """Get endpoint headers (auth + user custom). Override to customize.""" cfg = self.model_endpoint.endpoint - headers = dict(cfg.headers) if cfg.headers else {} + headers = cfg.get_headers(request_info.url_index) if cfg.api_key and not cfg.auth_type: headers["Authorization"] = f"Bearer {cfg.api_key}" return headers diff --git a/tests/unit/config/test_config_override_invariants.py b/tests/unit/config/test_config_override_invariants.py index 292abe3fb6..0ea27aa25d 100644 --- a/tests/unit/config/test_config_override_invariants.py +++ b/tests/unit/config/test_config_override_invariants.py @@ -97,6 +97,7 @@ def _unwrap(annotation: Any) -> Any: "request_rate": [2.0, 3.0], "benchmark_duration": [2.0, 3.0], "headers": [["x-probe-a:1"], ["x-probe-b:2"]], + "url_headers": [["0:x-probe-a:1"], ["0:x-probe-b:2"]], "extra_inputs": [["probe_a:1"], ["probe_b:2"]], "mlflow_tags": [["probe_a:1"], ["probe_b:2"]], "server_metrics": [ diff --git a/tests/unit/config/test_url_headers.py b/tests/unit/config/test_url_headers.py new file mode 100644 index 0000000000..16e8f7ad6a --- /dev/null +++ b/tests/unit/config/test_url_headers.py @@ -0,0 +1,221 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +"""Per-URL headers: parsing, CLI and YAML routing, and endpoint validation.""" + +from __future__ import annotations + +import textwrap +from pathlib import Path + +import orjson +import pytest +from cyclopts import App +from pydantic import ValidationError +from pytest import param + +from aiperf.common.models.model_endpoint_info import ModelEndpointInfo +from aiperf.config import BenchmarkRun +from aiperf.config.endpoint import EndpointConfig +from aiperf.config.flags._converter_endpoint import build_endpoint +from aiperf.config.flags.cli_config import CLIConfig +from aiperf.config.flags.resolver import resolve_config +from aiperf.config.loader import build_benchmark_plan, load_config_from_string +from aiperf.config.loader.parsing import parse_url_headers + +DP_URL = "http://localhost:8000" + + +def _parse_cli_args(argv: list[str]) -> CLIConfig: + captured: dict[str, CLIConfig] = {} + app = App(name="test_profile") + + @app.default + def _runner(*, cli_config: CLIConfig) -> None: + captured["config"] = cli_config + + try: + app(["--model", "test-model", *argv], exit_on_error=False) + except SystemExit as exc: + if exc.code not in (0, None): + raise + return captured["config"] + + +def _cli(**kwargs: object) -> CLIConfig: + """Build a CLIConfig whose model_fields_set is exactly ``kwargs``.""" + return CLIConfig(**CLIConfig(**kwargs).model_dump(exclude_unset=True)) # type: ignore[arg-type] + + +@pytest.mark.parametrize( + ("raw", "expected"), + [ + param(None, None, id="none"), + param("0:X-data-parallel-rank:0", [{"X-data-parallel-rank": "0"}], id="cli-form"), + param("2:X-Route:a", [{}, {}, {"X-Route": "a"}], id="sparse-index-pads"), + param(" 1 : X-Route : a:b:c ", [{}, {"X-Route": "a:b:c"}], id="value-keeps-colons"), + param('{"1": {"X-Route": "a"}}', [{}, {"X-Route": "a"}], id="json-index-mapping"), + param('[{}, {"X-Route": "a"}]', [{}, {"X-Route": "a"}], id="json-list"), + param([{"X-Route": 7}, {}], [{"X-Route": "7"}, {}], id="yaml-list-stringified"), + param({"1": {"X-Route": "b"}}, [{}, {"X-Route": "b"}], id="yaml-index-mapping"), + param( + ["0:X-Route:a", "0:X-Pool:p", "1:X-Route:b", "0:X-Route:c"], + [{"X-Route": "c", "X-Pool": "p"}, {"X-Route": "b"}], + id="repeated-flags-merge-last-wins", + ), + ], +) # fmt: skip +def test_parse_url_headers_accepted_forms(raw: object, expected: object) -> None: + assert parse_url_headers(raw) == expected + + +@pytest.mark.parametrize( + "raw", + [ + param("X-Route:a", id="missing-index"), + param("0:X-Route", id="missing-value"), + param("0::a", id="empty-name"), + param("-1:X-Route:a", id="negative-index"), + param("first:X-Route:a", id="non-integer-index"), + param({"x": {"X-Route": "a"}}, id="non-integer-mapping-key"), + param({"0": "X-Route:a"}, id="headers-not-a-mapping"), + param("[1, 2]", id="json-list-of-non-mappings"), + param("{not json", id="bad-json"), + param('"0:X:a"', id="quoted-index"), + param(3, id="unsupported-type"), + ], +) # fmt: skip +def test_parse_url_headers_invalid_raises_value_error(raw: object) -> None: + with pytest.raises(ValueError): + parse_url_headers(raw) + + +def test_url_headers_cli_flag_maps_to_endpoint() -> None: + config = _parse_cli_args( + [ + "--url", DP_URL, + "--url", DP_URL, + "--url-header", "0:X-data-parallel-rank:0", + "--url-header", "1:X-data-parallel-rank:1", + ] + ) # fmt: skip + + endpoint = build_endpoint(config) + + assert endpoint["urls"] == [DP_URL, DP_URL] + assert endpoint["url_headers"] == [ + {"X-data-parallel-rank": "0"}, + {"X-data-parallel-rank": "1"}, + ] + assert EndpointConfig(**endpoint).url_headers == endpoint["url_headers"] + + +def test_url_headers_unset_is_not_emitted_by_converter() -> None: + config = _parse_cli_args(["--url", DP_URL]) + + assert "url_headers" not in config.model_fields_set + assert "url_headers" not in build_endpoint(config) + + +def test_url_headers_index_out_of_range_raises() -> None: + with pytest.raises(ValidationError, match="URL index 2 is out of range"): + EndpointConfig(urls=[DP_URL, DP_URL], url_headers=["2:X-Route:a"]) + + +@pytest.mark.parametrize( + "name", + [ + param("Authorization", id="authorization"), + param("x-api-key", id="api-key"), + param("X-Tenant-Token", id="token-substring"), + ], +) # fmt: skip +def test_url_headers_credential_names_rejected(name: str) -> None: + with pytest.raises(ValidationError, match="credential-bearing"): + EndpointConfig(urls=[DP_URL], url_headers=[{name: "secret"}]) + + +def test_url_headers_survive_json_round_trip() -> None: + """Serialized runs reach worker processes as JSON.""" + original = EndpointConfig(urls=[DP_URL, DP_URL], url_headers=["1:X-Route:b"]) + + payload = orjson.loads(orjson.dumps(original.model_dump(mode="json"))) + restored = EndpointConfig.model_validate(payload) + + assert restored.url_headers == [{}, {"X-Route": "b"}] + + +def test_url_headers_yaml_reaches_runtime_endpoint() -> None: + config = load_config_from_string( + textwrap.dedent( + """\ + benchmark: + models: [test-model] + endpoint: + urls: [http://localhost:8000, http://localhost:8000] + type: chat + headers: + X-Team: perf + url_headers: + - {X-data-parallel-rank: "0"} + - {X-data-parallel-rank: "1"} + datasets: + - name: default + type: synthetic + phases: + - name: profiling + type: concurrency + concurrency: 1 + requests: 1 + """ + ), + substitute_env=True, + ) + plan = build_benchmark_plan(config) + assert len(plan.configs) == 1 + cfg = plan.configs[0] + run = BenchmarkRun( + benchmark_id="test-run", + sweep_id=plan.sweep_id, + cfg=cfg, + artifact_dir=cfg.artifacts.dir, + random_seed=None, + variables={}, + ) + + info = ModelEndpointInfo.from_run(run).endpoint + assert info.get_headers(0) == {"X-Team": "perf", "X-data-parallel-rank": "0"} + assert info.get_headers(1) == {"X-Team": "perf", "X-data-parallel-rank": "1"} + + +def test_url_headers_cli_overrides_config_file(tmp_path: Path) -> None: + config_file = tmp_path / "config.yaml" + config_file.write_text( + textwrap.dedent( + """\ + schemaVersion: "2.0" + benchmark: + model: test-model + endpoint: + urls: [http://localhost:8000, http://localhost:8000] + url_headers: + - {X-data-parallel-rank: "7"} + dataset: + type: synthetic + phases: + type: concurrency + concurrency: 1 + requests: 1 + """ + ) + ) + + resolved = resolve_config( + _cli(url_headers=["0:X-data-parallel-rank:0", "1:X-data-parallel-rank:1"]), + config_file, + ) + + assert resolved.benchmark.endpoint.url_headers == [ + {"X-data-parallel-rank": "0"}, + {"X-data-parallel-rank": "1"}, + ] + assert resolved.benchmark.endpoint.urls == [DP_URL, DP_URL] diff --git a/tests/unit/endpoints/test_endpoint_url_headers.py b/tests/unit/endpoints/test_endpoint_url_headers.py new file mode 100644 index 0000000000..552f3492a1 --- /dev/null +++ b/tests/unit/endpoints/test_endpoint_url_headers.py @@ -0,0 +1,100 @@ +# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. +# SPDX-License-Identifier: Apache-2.0 +"""Requests carry the headers of the URL they are sent to.""" + +from __future__ import annotations + +import pytest +from pytest import param + +from aiperf.common.enums import ModelSelectionStrategy +from aiperf.common.models.model_endpoint_info import ( + EndpointInfo, + ModelEndpointInfo, + ModelInfo, + ModelListInfo, +) +from aiperf.endpoints.anthropic_messages import MessagesEndpoint +from aiperf.endpoints.openai_chat import ChatEndpoint +from aiperf.plugin.enums import EndpointType +from tests.unit.endpoints.conftest import ( + create_endpoint_with_mock_transport, + create_request_info, +) + +DP_URL = "http://localhost:8000" + + +def _model_endpoint( + endpoint_type: EndpointType = EndpointType.CHAT, + api_key: str | None = None, +) -> ModelEndpointInfo: + return ModelEndpointInfo( + models=ModelListInfo( + models=[ModelInfo(name="test-model")], + model_selection_strategy=ModelSelectionStrategy.ROUND_ROBIN, + ), + endpoint=EndpointInfo( + type=endpoint_type, + base_urls=[DP_URL, DP_URL, DP_URL], + headers=[("X-Team", "perf"), ("x-data-parallel-rank", "global")], + url_headers=[ + {"X-data-parallel-rank": "0"}, + {}, + {"X-data-parallel-rank": "2", "X-Pool": "b"}, + ], + api_key=api_key, + ), + ) + + +@pytest.mark.parametrize( + ("url_index", "expected"), + [ + param(None, {"X-Team": "perf", "X-data-parallel-rank": "0"}, id="unset-index-uses-first-url"), + param(0, {"X-Team": "perf", "X-data-parallel-rank": "0"}, id="url-0"), + param(1, {"X-Team": "perf", "x-data-parallel-rank": "global"}, id="url-without-overrides"), + param(2, {"X-Team": "perf", "X-data-parallel-rank": "2", "X-Pool": "b"}, id="url-2"), + param(3, {"X-Team": "perf", "X-data-parallel-rank": "0"}, id="index-wraps-like-get-url"), + ], +) # fmt: skip +def test_get_headers_overlays_url_headers_case_insensitively( + url_index: int | None, expected: dict[str, str] +) -> None: + assert _model_endpoint().endpoint.get_headers(url_index) == expected + + +def test_get_headers_without_url_headers_returns_global_headers() -> None: + info = EndpointInfo(base_urls=[DP_URL, DP_URL], headers=[("X-Team", "perf")]) + + assert info.get_headers(0) == info.get_headers(1) == {"X-Team": "perf"} + + +@pytest.mark.parametrize("url_index", [0, 1, 2]) +def test_chat_endpoint_headers_follow_request_url_index(url_index: int) -> None: + model_endpoint = _model_endpoint(api_key="sk-test") + endpoint = create_endpoint_with_mock_transport(ChatEndpoint, model_endpoint) + request_info = create_request_info(model_endpoint=model_endpoint) + request_info.url_index = url_index + + headers = endpoint.get_endpoint_headers(request_info) + + assert headers == { + **model_endpoint.endpoint.get_headers(url_index), + "Authorization": "Bearer sk-test", + } + + +def test_messages_endpoint_headers_follow_request_url_index() -> None: + model_endpoint = _model_endpoint(EndpointType.MESSAGES, api_key="sk-ant") + endpoint = create_endpoint_with_mock_transport(MessagesEndpoint, model_endpoint) + request_info = create_request_info(model_endpoint=model_endpoint) + request_info.url_index = 2 + + headers = endpoint.get_endpoint_headers(request_info) + + assert headers["X-data-parallel-rank"] == "2" + assert headers["X-Pool"] == "b" + assert "x-data-parallel-rank" not in headers + assert headers["x-api-key"] == "sk-ant" + assert headers["anthropic-version"] == "2023-06-01" diff --git a/tools/ergonomics_baseline.json b/tools/ergonomics_baseline.json index 570e734af2..739cb1db1c 100644 --- a/tools/ergonomics_baseline.json +++ b/tools/ergonomics_baseline.json @@ -396,6 +396,11 @@ "src/aiperf/common/models/export_models.py", "JsonExportData" ], + [ + "pydantic-fields", + "src/aiperf/config/endpoint.py", + "EndpointConfig" + ], [ "pydantic-fields", "src/aiperf/kubernetes/environment.py", From 39fc5baca6e5aaed8ca94fac9ea4416e32d31e6d Mon Sep 17 00:00:00 2001 From: Dundy Pasupuleti Date: Wed, 11 Nov 2026 10:55:03 -0500 Subject: [PATCH 2/4] fix(endpoint): bound url header indices before padding, replace names case-insensitively Check a per-URL header index against the configured URL count before any padding is allocated, so an index like 1000000000 fails with the out of range error instead of building a billion empty mappings. When the URL count is not known yet (the CLI flag, which may be overlaid on a config file's urls), indices are capped at MAX_URL_HEADER_INDEX. Repeated header names for one URL now replace each other case-insensitively, so 0:X-Route:a followed by 0:x-route:b sends only the later value. Signed-off-by: Dundy Pasupuleti --- src/aiperf/config/endpoint.py | 14 +++-- src/aiperf/config/loader/parsing.py | 84 ++++++++++++++++++++++----- tests/unit/config/test_url_headers.py | 29 +++++++++ 3 files changed, 106 insertions(+), 21 deletions(-) diff --git a/src/aiperf/config/endpoint.py b/src/aiperf/config/endpoint.py index 814990877c..b77d862008 100644 --- a/src/aiperf/config/endpoint.py +++ b/src/aiperf/config/endpoint.py @@ -38,7 +38,11 @@ parse_enabled_or_config, require_relative_path, ) -from aiperf.config.loader.parsing import normalize_http_urls, parse_url_headers +from aiperf.config.loader.parsing import ( + normalize_http_urls, + parse_url_headers_for_urls, + url_index_out_of_range_message, +) from aiperf.config.sagemaker import SageMakerConfig from aiperf.plugin.enums import ( EndpointType, @@ -559,7 +563,7 @@ def _redact_headers(self, value: dict[str, str]) -> dict[str, str]: "``headers`` or ``api_key``. " "Example: [{X-data-parallel-rank: '0'}, {X-data-parallel-rank: '1'}]", ), - BeforeValidator(parse_url_headers), + BeforeValidator(parse_url_headers_for_urls), ] extra: Annotated[ @@ -859,9 +863,9 @@ def _validate_url_headers(self) -> Self: if len(self.url_headers) > len(self.urls): raise ValueError( - f"url_headers: URL index {len(self.url_headers) - 1} is out of " - f"range; {len(self.urls)} URL(s) are configured, so valid " - f"indices are 0..{len(self.urls) - 1}." + url_index_out_of_range_message( + len(self.url_headers) - 1, len(self.urls) + ) ) for index, headers in enumerate(self.url_headers): sensitive = sorted(n for n in headers if is_sensitive_header_name(n)) diff --git a/src/aiperf/config/loader/parsing.py b/src/aiperf/config/loader/parsing.py index 01a3624441..221aa086d9 100644 --- a/src/aiperf/config/loader/parsing.py +++ b/src/aiperf/config/loader/parsing.py @@ -6,6 +6,7 @@ from typing import TYPE_CHECKING, Any import orjson +from pydantic import ValidationInfo from aiperf.common.utils import load_json_str from aiperf.plugin.enums import ServiceType @@ -242,6 +243,10 @@ def parse_str_or_dict_as_tuple_list(input: Any | None) -> list[tuple[str, Any]] raise ValueError(f"User Config: {input} - must be a valid string, list, or dict") +MAX_URL_HEADER_INDEX = 1023 +"""Upper bound for a per-URL header index when the URL count is not known.""" + + def _parse_url_index(raw: Any, source: Any) -> int: """Parse a per-URL header index (the 0-based position in the URL list).""" if isinstance(raw, int) and not isinstance(raw, bool) and raw >= 0: @@ -253,16 +258,35 @@ def _parse_url_index(raw: Any, source: Any) -> int: ) +def url_index_out_of_range_message(index: int, url_count: int) -> str: + """Error text for a per-URL header index past the configured URLs.""" + return ( + f"url_headers: URL index {index} is out of range; {url_count} URL(s) " + f"are configured, so valid indices are 0..{url_count - 1}." + ) + + def _set_url_headers( - output: list[dict[str, str]], index: int, headers: Any, source: Any + output: dict[int, dict[str, str]], + index: int, + headers: Any, + source: Any, + *, + url_count: int | None, ) -> None: + if url_count is not None and index >= url_count: + raise ValueError(url_index_out_of_range_message(index, url_count=url_count)) + if index > MAX_URL_HEADER_INDEX: + raise ValueError( + f"User Config: {source!r} - URL index {index} is above the " + f"maximum of {MAX_URL_HEADER_INDEX}" + ) if not isinstance(headers, dict): raise ValueError( f"User Config: {source!r} - headers for URL {index} must be a " "mapping of header name to value" ) - while len(output) <= index: - output.append({}) + target = output.setdefault(index, {}) for name, value in headers.items(): name = str(name).strip() if not name: @@ -270,10 +294,16 @@ def _set_url_headers( f"User Config: {source!r} - header name for URL {index} " "must not be empty" ) - output[index][name] = str(value) + # Header names are case-insensitive: a later assignment replaces an + # earlier one regardless of case instead of sending both. + for existing in [k for k in target if k.lower() == name.lower()]: + del target[existing] + target[name] = str(value) -def _merge_url_headers_item(output: list[dict[str, str]], item: Any) -> None: +def _merge_url_headers_item( + output: dict[int, dict[str, str]], item: Any, url_count: int | None +) -> None: if isinstance(item, str) and item.lstrip().startswith(("{", "[")): try: item = load_json_str(item) @@ -283,11 +313,17 @@ def _merge_url_headers_item(output: list[dict[str, str]], item: Any) -> None: ) from e if isinstance(item, list): for index, headers in enumerate(item): - _set_url_headers(output, index, headers, item) + _set_url_headers(output, index, headers, item, url_count=url_count) return if isinstance(item, dict): for raw_index, headers in item.items(): - _set_url_headers(output, _parse_url_index(raw_index, item), headers, item) + _set_url_headers( + output, + _parse_url_index(raw_index, item), + headers, + item, + url_count=url_count, + ) elif isinstance(item, str): parts = item.split(":", 2) if len(parts) != 3: @@ -296,7 +332,9 @@ def _merge_url_headers_item(output: list[dict[str, str]], item: Any) -> None: "format, e.g. '0:X-data-parallel-rank:0'" ) index = _parse_url_index(parts[0], item) - _set_url_headers(output, index, {parts[1]: parts[2].strip()}, item) + _set_url_headers( + output, index, {parts[1]: parts[2].strip()}, item, url_count=url_count + ) else: raise ValueError( f"User Config: {item!r} - must be an 'INDEX:Header:Value' string, " @@ -304,7 +342,9 @@ def _merge_url_headers_item(output: list[dict[str, str]], item: Any) -> None: ) -def parse_url_headers(input: Any) -> list[dict[str, str]] | None: +def parse_url_headers( + input: Any, url_count: int | None = None +) -> list[dict[str, str]] | None: """Parse per-URL headers into a list aligned with the URL list. Entry ``i`` holds the headers for the ``i``-th URL; URLs without headers @@ -313,18 +353,30 @@ def parse_url_headers(input: Any) -> list[dict[str, str]] | None: of URL index to headers, a JSON string of either, or a list of CLI strings. Later entries for the same header win. Values are kept as strings: unlike ``--header`` they are never coerced, so - ``0:X-data-parallel-rank:0`` sends ``"0"``. + ``0:X-data-parallel-rank:0`` sends ``"0"``. Header names are matched + case-insensitively, so a later ``x-route`` replaces an earlier ``X-Route``. + + Indices are checked against ``url_count`` (when known) before any padding + is allocated; without it they are capped at ``MAX_URL_HEADER_INDEX``. """ if input is None: return None - output: list[dict[str, str]] = [] + output: dict[int, dict[str, str]] = {} if isinstance(input, (list, tuple)) and all(isinstance(i, dict) for i in input): for index, headers in enumerate(input): - _set_url_headers(output, index, headers, input) - return output - for item in input if isinstance(input, (list, tuple)) else [input]: - _merge_url_headers_item(output, item) - return output + _set_url_headers(output, index, headers, input, url_count=url_count) + else: + for item in input if isinstance(input, (list, tuple)) else [input]: + _merge_url_headers_item(output, item, url_count) + size = max(output) + 1 if output else 0 + return [output.get(index, {}) for index in range(size)] + + +def parse_url_headers_for_urls(input: Any, info: ValidationInfo) -> Any: + """``parse_url_headers`` bound to the already-validated ``urls`` field.""" + urls = info.data.get("urls") + url_count = len(urls) if isinstance(urls, (list, tuple)) else None + return parse_url_headers(input, url_count) def print_str_or_list(input: Any) -> str: diff --git a/tests/unit/config/test_url_headers.py b/tests/unit/config/test_url_headers.py index 16e8f7ad6a..6965a519e9 100644 --- a/tests/unit/config/test_url_headers.py +++ b/tests/unit/config/test_url_headers.py @@ -62,6 +62,16 @@ def _cli(**kwargs: object) -> CLIConfig: [{"X-Route": "c", "X-Pool": "p"}, {"X-Route": "b"}], id="repeated-flags-merge-last-wins", ), + param( + ["0:X-Route:a", "0:x-route:b", "0:X-ROUTE:c"], + [{"X-ROUTE": "c"}], + id="repeated-names-replace-case-insensitively", + ), + param( + {"0": {"X-Route": "a", "x-route": "b"}}, + [{"x-route": "b"}], + id="mapping-names-replace-case-insensitively", + ), ], ) # fmt: skip def test_parse_url_headers_accepted_forms(raw: object, expected: object) -> None: @@ -82,6 +92,7 @@ def test_parse_url_headers_accepted_forms(raw: object, expected: object) -> None param("{not json", id="bad-json"), param('"0:X:a"', id="quoted-index"), param(3, id="unsupported-type"), + param("1000000000:X-Route:a", id="index-above-cap"), ], ) # fmt: skip def test_parse_url_headers_invalid_raises_value_error(raw: object) -> None: @@ -121,6 +132,24 @@ def test_url_headers_index_out_of_range_raises() -> None: EndpointConfig(urls=[DP_URL, DP_URL], url_headers=["2:X-Route:a"]) +@pytest.mark.parametrize( + "url_headers", + [ + param(["1000000000:X-Route:a"], id="cli-form"), + param({"1000000000": {"X-Route": "a"}}, id="index-mapping"), + ], +) # fmt: skip +def test_url_headers_huge_index_is_rejected_before_padding(url_headers: object) -> None: + with pytest.raises(ValidationError, match="URL index 1000000000 is out of range"): + EndpointConfig(urls=[DP_URL, DP_URL], url_headers=url_headers) + + +def test_parse_url_headers_checks_url_count_before_padding() -> None: + with pytest.raises(ValueError, match="URL index 5 is out of range"): + parse_url_headers("5:X-Route:a", url_count=2) + assert parse_url_headers("1:X-Route:a", url_count=2) == [{}, {"X-Route": "a"}] + + @pytest.mark.parametrize( "name", [ From fe91c753831191348e05e82abc0767370c80126b Mon Sep 17 00:00:00 2001 From: Dundy Pasupuleti Date: Wed, 11 Nov 2026 11:53:42 -0500 Subject: [PATCH 3/4] fix(config): apply the fixed url header index cap only when the url count is unknown Signed-off-by: Dundy Pasupuleti --- src/aiperf/config/loader/parsing.py | 2 +- tests/unit/config/test_url_headers.py | 19 ++++++++++++++++++- 2 files changed, 19 insertions(+), 2 deletions(-) diff --git a/src/aiperf/config/loader/parsing.py b/src/aiperf/config/loader/parsing.py index 221aa086d9..0794068ef8 100644 --- a/src/aiperf/config/loader/parsing.py +++ b/src/aiperf/config/loader/parsing.py @@ -276,7 +276,7 @@ def _set_url_headers( ) -> None: if url_count is not None and index >= url_count: raise ValueError(url_index_out_of_range_message(index, url_count=url_count)) - if index > MAX_URL_HEADER_INDEX: + if url_count is None and index > MAX_URL_HEADER_INDEX: raise ValueError( f"User Config: {source!r} - URL index {index} is above the " f"maximum of {MAX_URL_HEADER_INDEX}" diff --git a/tests/unit/config/test_url_headers.py b/tests/unit/config/test_url_headers.py index 6965a519e9..1afeff48dc 100644 --- a/tests/unit/config/test_url_headers.py +++ b/tests/unit/config/test_url_headers.py @@ -20,7 +20,7 @@ from aiperf.config.flags.cli_config import CLIConfig from aiperf.config.flags.resolver import resolve_config from aiperf.config.loader import build_benchmark_plan, load_config_from_string -from aiperf.config.loader.parsing import parse_url_headers +from aiperf.config.loader.parsing import MAX_URL_HEADER_INDEX, parse_url_headers DP_URL = "http://localhost:8000" @@ -150,6 +150,23 @@ def test_parse_url_headers_checks_url_count_before_padding() -> None: assert parse_url_headers("1:X-Route:a", url_count=2) == [{}, {"X-Route": "a"}] +def test_parse_url_headers_known_url_count_is_the_only_bound() -> None: + # with the url count known, the fixed cap does not apply on top of it + url_count = MAX_URL_HEADER_INDEX + 2 + aligned = [{} for _ in range(url_count - 1)] + [{"X-Route": "last"}] + parsed = parse_url_headers(aligned, url_count=url_count) + assert parsed is not None + assert len(parsed) == url_count + assert parsed[-1] == {"X-Route": "last"} + last = url_count - 1 + assert parse_url_headers(f"{last}:X-Route:a", url_count=url_count)[-1] == { + "X-Route": "a" + } + # without a url count the fixed cap still guards padding + with pytest.raises(ValueError, match="maximum"): + parse_url_headers(f"{MAX_URL_HEADER_INDEX + 1}:X-Route:a") + + @pytest.mark.parametrize( "name", [ From 53f9c51b8bd8f9cfe17e8dfdb27a807be1598163 Mon Sep 17 00:00:00 2001 From: Dundy Pasupuleti Date: Wed, 11 Nov 2026 12:57:30 -0500 Subject: [PATCH 4/4] test(config): share the exact-fields-set CLIConfig helper via tests.unit.conftest Signed-off-by: Dundy Pasupuleti --- tests/unit/config/test_raw_implies_outputs_json.py | 14 ++++++-------- tests/unit/config/test_url_headers.py | 10 ++++------ tests/unit/conftest.py | 5 +++++ 3 files changed, 15 insertions(+), 14 deletions(-) diff --git a/tests/unit/config/test_raw_implies_outputs_json.py b/tests/unit/config/test_raw_implies_outputs_json.py index 058e0d63de..6d696f3e23 100644 --- a/tests/unit/config/test_raw_implies_outputs_json.py +++ b/tests/unit/config/test_raw_implies_outputs_json.py @@ -19,6 +19,7 @@ from aiperf.config.flags.resolver import resolve_config from aiperf.config.loader import load_config_from_mapping from aiperf.kubernetes.spec_converter import build_config_envelope +from tests.unit.conftest import make_cli_with_fields_set def _artifacts(**kwargs: Any) -> ArtifactsConfig: @@ -103,20 +104,17 @@ def _yaml(tmp_path: Path, artifacts: str) -> Path: ) return cfg - @staticmethod - def _cli(**kwargs: Any) -> CLIConfig: - """A CLIConfig whose model_fields_set is exactly ``kwargs``.""" - return CLIConfig(**CLIConfig(**kwargs).model_dump(exclude_unset=True)) - def test_yaml_raw_true_implies_outputs_json(self, tmp_path: Path) -> None: - cfg = resolve_config(self._cli(), self._yaml(tmp_path, " raw: true\n")) + cfg = resolve_config( + make_cli_with_fields_set(), self._yaml(tmp_path, " raw: true\n") + ) assert cfg.benchmark.artifacts.export_outputs_json is True def test_yaml_explicit_false_survives_cli_export_level_raw( self, tmp_path: Path ) -> None: cfg = resolve_config( - self._cli(export_level=ExportLevel.RAW), + make_cli_with_fields_set(export_level=ExportLevel.RAW), self._yaml(tmp_path, " exportOutputsJson: false\n"), ) artifacts = cfg.benchmark.artifacts @@ -125,7 +123,7 @@ def test_yaml_explicit_false_survives_cli_export_level_raw( def test_yaml_raw_true_with_cli_opt_out(self, tmp_path: Path) -> None: cfg = resolve_config( - self._cli(export_outputs_json=False), + make_cli_with_fields_set(export_outputs_json=False), self._yaml(tmp_path, " raw: true\n"), ) artifacts = cfg.benchmark.artifacts diff --git a/tests/unit/config/test_url_headers.py b/tests/unit/config/test_url_headers.py index 1afeff48dc..3ae6a6f510 100644 --- a/tests/unit/config/test_url_headers.py +++ b/tests/unit/config/test_url_headers.py @@ -21,6 +21,7 @@ from aiperf.config.flags.resolver import resolve_config from aiperf.config.loader import build_benchmark_plan, load_config_from_string from aiperf.config.loader.parsing import MAX_URL_HEADER_INDEX, parse_url_headers +from tests.unit.conftest import make_cli_with_fields_set DP_URL = "http://localhost:8000" @@ -41,11 +42,6 @@ def _runner(*, cli_config: CLIConfig) -> None: return captured["config"] -def _cli(**kwargs: object) -> CLIConfig: - """Build a CLIConfig whose model_fields_set is exactly ``kwargs``.""" - return CLIConfig(**CLIConfig(**kwargs).model_dump(exclude_unset=True)) # type: ignore[arg-type] - - @pytest.mark.parametrize( ("raw", "expected"), [ @@ -256,7 +252,9 @@ def test_url_headers_cli_overrides_config_file(tmp_path: Path) -> None: ) resolved = resolve_config( - _cli(url_headers=["0:X-data-parallel-rank:0", "1:X-data-parallel-rank:1"]), + make_cli_with_fields_set( + url_headers=["0:X-data-parallel-rank:0", "1:X-data-parallel-rank:1"] + ), config_file, ) diff --git a/tests/unit/conftest.py b/tests/unit/conftest.py index ca74a087e9..b9bda8560a 100644 --- a/tests/unit/conftest.py +++ b/tests/unit/conftest.py @@ -589,6 +589,11 @@ def cli_config() -> CLIConfig: return CLIConfig(model_names=["test-model"]) +def make_cli_with_fields_set(**kwargs: object) -> CLIConfig: + """Build a CLIConfig whose model_fields_set is exactly ``kwargs``.""" + return CLIConfig(**CLIConfig(**kwargs).model_dump(exclude_unset=True)) # type: ignore[arg-type] + + def make_run_from_cli( cli_config: CLIConfig, ):