vllm_dissag: remove tests/ scaffolding + follow-ups on #171

scripts/vllm_dissag/ARCHITECTURE.md

@@ -12,31 +12,31 @@ How the pieces fit, from sbatch down to the per-node `vllm serve` workers.
 
 ```mermaid
 flowchart TB
-    user([user / CI]) -->|"sbatch + env\n(CONNECTOR, WIDE_EP, EP_BACKEND,\nMODEL_NAME, xP, yD, RUN_MORI/RUN_DEEPEP)"| slurm
+    user([user / CI]) -->|"sbatch + env<br/>(CONNECTOR, WIDE_EP, EP_BACKEND,<br/>MODEL_NAME, xP, yD, RUN_MORI/RUN_DEEPEP)"| slurm
 
     subgraph host["Submit host"]
-        slurm["run_xPyD_models.slurm\n• resolve script dir\n• validate MODEL_NAME (VALID_MODELS)\n• axis resolution + back-compat shim\n• pick nodes, gather IPs\n• docker run per node (-e env plumb)"]
+        slurm["run_xPyD_models.slurm<br/>• resolve script dir<br/>• validate MODEL_NAME (VALID_MODELS)<br/>• axis resolution + back-compat shim<br/>• pick nodes, gather IPs<br/>• docker run per node (-e env plumb)"]
     end
 
-    slurm -->|"srun + docker run\n(one container per node)"| n0
+    slurm -->|"srun + docker run<br/>(one container per node)"| n0
     slurm --> n1
     slurm --> n2
     slurm --> n3
 
     subgraph cluster["Allocated nodes (xP prefill + yD decode)"]
-        n0["NODE_RANK 0\nPrefill MASTER + Proxy"]
-        n1["NODE_RANK 1..xP-1\nPrefill CHILD"]
-        n2["NODE_RANK xP\nDecode MASTER"]
-        n3["NODE_RANK xP+1..end\nDecode CHILD"]
+        n0["NODE_RANK 0<br/>Prefill MASTER + Proxy"]
+        n1["NODE_RANK 1..xP-1<br/>Prefill CHILD"]
+        n2["NODE_RANK xP<br/>Decode MASTER"]
+        n3["NODE_RANK xP+1..end<br/>Decode CHILD"]
     end
 
     subgraph driverbox["Inside each container: vllm_disagg.sh (the one launcher)"]
-        driver["DRIVER\n• axis select + validate\n• topology math\n• models.yaml parse\n• role branch\n• barrier / benchmark / cleanup"]
-        para["parallelism.sh\nTP vs wideEP arg helpers"]
-        yaml[("models.yaml\nper-model flags + env")]
+        driver["DRIVER<br/>• axis select + validate<br/>• topology math<br/>• models.yaml parse<br/>• role branch<br/>• barrier / benchmark / cleanup"]
+        para["parallelism.sh<br/>TP vs wideEP arg helpers"]
+        yaml[("models.yaml<br/>per-model flags + env")]
         conn{{"connectors/<CONNECTOR>.sh"}}
-        rixl["rixl.sh\nNixlConnector\nTP + DeepEP"]
-        moriio["moriio.sh\nMoRIIOConnector\nMoRIIO+TP + MoRI-EP"]
+        rixl["rixl.sh<br/>NixlConnector<br/>TP + DeepEP"]
+        moriio["moriio.sh<br/>MoRIIOConnector<br/>MoRIIO+TP + MoRI-EP"]
         driver --> para
         driver --> yaml
         driver --> conn
@@ -45,9 +45,9 @@ flowchart TB
     end
 
     n0 --> driverbox
-    driver -->|connector_launch_worker| vllm[["vllm serve\n(prefill / decode worker)"]]
-    driver -->|rank 0 only| proxy[["proxy / router\n(co-located)"]]
-    proxy --> bench["benchmark_xPyD.sh\n→ *_CONCURRENCY.log"]
+    driver -->|connector_launch_worker| vllm[["vllm serve<br/>(prefill / decode worker)"]]
+    driver -->|rank 0 only| proxy[["proxy / router<br/>(co-located)"]]
+    proxy --> bench["benchmark_xPyD.sh<br/>→ *_CONCURRENCY.log"]
 ```
 
 ---
@@ -62,24 +62,24 @@ flowchart TD
     start([env in]) --> q0{CONNECTOR set?}
 
     q0 -->|no| shim{legacy flag?}
-    shim -->|RUN_MORI=1| m["CONNECTOR=moriio\nWIDE_EP=1\nEP_BACKEND=mori"]
-    shim -->|RUN_DEEPEP=1| d["CONNECTOR=rixl\nWIDE_EP=1\nEP_BACKEND=deepep"]
-    shim -->|neither| def["CONNECTOR=rixl\nWIDE_EP=0  (TP)"]
-    q0 -->|yes| explicit["use explicit\nCONNECTOR / WIDE_EP / EP_BACKEND"]
+    shim -->|RUN_MORI=1| m["CONNECTOR=moriio<br/>WIDE_EP=1<br/>EP_BACKEND=mori"]
+    shim -->|RUN_DEEPEP=1| d["CONNECTOR=rixl<br/>WIDE_EP=1<br/>EP_BACKEND=deepep"]
+    shim -->|neither| def["CONNECTOR=rixl<br/>WIDE_EP=0  (TP)"]
+    q0 -->|yes| explicit["use explicit<br/>CONNECTOR / WIDE_EP / EP_BACKEND"]
 
     m --> vconn
     d --> vconn
     def --> vconn
     explicit --> vconn
 
-    vconn{validate CONNECTOR\nin rixl|moriio}
+    vconn{"validate CONNECTOR<br/>in rixl|moriio"}
     vconn -->|invalid| err1[["abort: invalid CONNECTOR"]]
-    vconn -->|valid| vwide{validate WIDE_EP\nin 0|1}
+    vconn -->|valid| vwide{"validate WIDE_EP<br/>in 0|1"}
     vwide -->|invalid| err2[["abort: invalid WIDE_EP"]]
     vwide -->|valid| qwide{WIDE_EP == 1?}
 
-    qwide -->|no  (TP)| okTP["EP_BACKEND = n/a"]
-    qwide -->|yes wideEP| qpair{connector ↔ EP_BACKEND}
+    qwide -->|"no (TP)"| okTP["EP_BACKEND = n/a"]
+    qwide -->|yes wideEP| qpair{"connector ↔ EP_BACKEND"}
     qpair -->|moriio + mori| okM["OK: all2all = mori_*"]
     qpair -->|rixl + deepep| okD["OK: all2all = deepep_*"]
     qpair -->|moriio + deepep| err3[["abort: cross-pair"]]
@@ -95,12 +95,12 @@ flowchart TD
 ```mermaid
 flowchart LR
     subgraph TP["WIDE_EP=0  (TP, PARALLEL_MODE=tp)"]
-        rt["rixl + TP\nNixlConnector\n--tensor-parallel-size"]
-        mt["moriio + TP  (NEW)\nMoRIIOConnector\n--tensor-parallel-size"]
+        rt["rixl + TP<br/>NixlConnector<br/>--tensor-parallel-size"]
+        mt["moriio + TP  (NEW)<br/>MoRIIOConnector<br/>--tensor-parallel-size"]
     end
     subgraph EP["WIDE_EP=1  (wideEP, PARALLEL_MODE=dp)"]
-        rd["rixl + deepep\nNixlConnector\n-tp 1 --data-parallel-size\n--enable-expert-parallel\n--all2all-backend deepep_*"]
-        md["moriio + mori\nMoRIIOConnector\n-tp 1 --data-parallel-size\n--enable-expert-parallel\n--all2all-backend mori_*"]
+        rd["rixl + deepep<br/>NixlConnector<br/>-tp 1 --data-parallel-size<br/>--enable-expert-parallel<br/>--all2all-backend deepep_*"]
+        md["moriio + mori<br/>MoRIIOConnector<br/>-tp 1 --data-parallel-size<br/>--enable-expert-parallel<br/>--all2all-backend mori_*"]
     end
 ```
 
@@ -116,16 +116,16 @@ stateDiagram-v2
     [*] --> ResolveAxes
     ResolveAxes --> ParseModel: CONNECTOR/WIDE_EP/EP_BACKEND valid
     ResolveAxes --> Abort: invalid / cross-pair
-    ParseModel: Parse models.yaml\n(export env: block,\nresolve prefill/decode flags)
+    ParseModel: Parse models.yaml<br/>(export env: block,<br/>resolve prefill/decode flags)
     ParseModel --> LoadProfiles
-    LoadProfiles: source parallelism.sh +\nconnectors/<CONNECTOR>.sh →\nconnector_init
+    LoadProfiles: source parallelism.sh +<br/>connectors/<CONNECTOR>.sh →<br/>connector_init
 
     LoadProfiles --> DryRunEmit: DRY_RUN=1
-    DryRunEmit: emit assembled\nvllm serve argv
+    DryRunEmit: emit assembled<br/>vllm serve argv
     DryRunEmit --> [*]
 
     LoadProfiles --> Barrier: normal run
-    Barrier: container barrier\n(socket_barrier.py)\n+ connector_runtime_patch
+    Barrier: container barrier<br/>(socket_barrier.py)<br/>+ connector_runtime_patch
     Barrier --> RoleBranch
 
     state RoleBranch <<choice>>
@@ -134,16 +134,16 @@ stateDiagram-v2
     RoleBranch --> DecodeMaster: RANK==xP
     RoleBranch --> DecodeChild: RANK>xP
 
-    PrefillMaster: launch worker (kv_producer)\n+ wait_workers_ready\n+ start_proxy + benchmark
+    PrefillMaster: launch worker (kv_producer)<br/>+ wait_workers_ready<br/>+ start_proxy + benchmark
     PrefillChild: launch worker (kv_producer, headless)
     DecodeMaster: launch worker (kv_consumer)
     DecodeChild: launch worker (kv_consumer, headless)
 
-    PrefillMaster --> Cleanup: bench done →\nkill proxy + worker
+    PrefillMaster --> Cleanup: bench done →<br/>kill proxy + worker
     PrefillChild --> WaitProxy
     DecodeMaster --> WaitProxy
     DecodeChild --> WaitProxy
-    WaitProxy: wait_for_proxy_and_cleanup\n(barrier → wait close → kill)
+    WaitProxy: wait_for_proxy_and_cleanup<br/>(barrier → wait close → kill)
     WaitProxy --> Cleanup
     Cleanup --> [*]
     Abort --> [*]
@@ -189,12 +189,12 @@ sequenceDiagram
 
 | Hook | rixl.sh | moriio.sh |
 |------|---------|-----------|
-| `connector_init` | ports 2584/router; PROXY_TYPE; barrier 5000/15000 | ports 20005/10001; per-role MoRI backend; barrier 2222 |
+| `connector_init` | serve 2584; `ROUTER_PORT` 18001; barrier 5000 (TP) / 15000 (deepep); `PROXY_TYPE` vllm_router\|toy_proxy | serve 20005; `ROUTER_PORT` 30000, toy 10001; barrier 2222; `PROXY_TYPE` vllm_router\|moriio_toy |
 | `connector_setup_env` | UCX/NIXL (TP) or RocSHMEM/UCX/NIXL+#39276 sed (deepep) | MoRI/RDMA/RocSHMEM + caches |
 | `connector_runtime_patch` | no-op (deepep patches inline in setup_env) | no-op (disagg fixes are in-source in the image's vLLM) |
 | `connector_launch_worker` | TP server **or** deepep DP+EP server | MoRIIO+TP server **or** MoRI-EP DP+EP server |
 | `connector_wait_workers_ready` | TP: socket_barrier; deepep: log-signal | log-signal on prefill+decode masters |
-| `connector_start_proxy` | vllm_router / toy_proxy over all P/D IPs | moriio_toy_proxy (or router) + curl probe |
+| `connector_start_proxy` | vllm_router / toy_proxy over all P/D IPs | vllm_router / moriio_toy + curl probe |
 
 ---
 
@@ -206,19 +206,19 @@ Where each piece of configuration lives, and the precedence when they overlap.
 flowchart TD
     subgraph yamlcat["models.yaml (per model)"]
         bf["base_flags"]
-        mf["tp_flags / dp_flags\n(by WIDE_EP)"]
+        mf["tp_flags / dp_flags<br/>(by WIDE_EP)"]
         rf["prefill.{tp,dp} / decode.{tp,dp}"]
         ef["experimental_flags"]
         envb["env: { VAR: val }"]
     end
-    bf & mf & rf & ef --> compose["MODEL_CONFIG_PREFILL / _DECODE\n(passed to connector_launch_worker)"]
+    bf & mf & rf & ef --> compose["MODEL_CONFIG_PREFILL / _DECODE<br/>(passed to connector_launch_worker)"]
 
     subgraph envlayer["ENV precedence (low → high)"]
-        c1["connector default\nexport VAR=\${VAR:-default}"] --> c2["models.yaml env:\n(exported before setup_env)"] --> c3["slurm -e VAR=…\n(submit/site override, wins)"]
+        c1["connector default<br/>export VAR=\${VAR:-default}"] --> c2["models.yaml env:<br/>(exported before setup_env)"] --> c3["slurm -e VAR=…<br/>(submit/site override, wins)"]
     end
     envb -.-> c2
 
-    note["Launcher owns: connector, transfer,\nparallelism DEGREE (--tensor/data-parallel-size,\n--all2all-backend, kv-transfer-config).\nyaml owns: model-tuning flags + env only."]
+    note["Launcher owns: connector, transfer,<br/>parallelism DEGREE (--tensor/data-parallel-size,<br/>--all2all-backend, kv-transfer-config).<br/>yaml owns: model-tuning flags + env only."]
 ```
 
 ---
@@ -234,10 +234,5 @@ flowchart TD
 | `connectors/moriio.sh` | MoRIIOConnector profile (MoRIIO+TP + MoRI-EP) |
 | `connectors/{rixl,moriio}.env` | per-connector platform env (expandable_segments:False etc.), forwarded via `docker -e` |
 | `models.yaml` | per-model flags + env catalog |
-| `tests/gate_check.sh` | combo-gate unit tests (model × connector × WIDE_EP allow/reject) |
-| `tests/argv_assert.sh` | per-cell `vllm serve` flag/env assertions from the driver's `DRY_RUN=1` output |
-| `tests/run_all.sh` | runs all offline gates (gate_check + argv_assert) |
-| `tests/{drive_cell,harvest,run_interactive}.sh` | interactive-allocation live-test drivers |
-| `tests/TEST_PLAN.md` | before/after verification plan |
 | `benchmark_xPyD.sh`, `benchmark_long_context.sh`, `benchmark_niah.{sh,py}`, `benchmark_parser.py`, `parse_to_csv.py` | benchmark + parsing (NIAH = long-context retrieval, vllm#47042) |
 | `socket_barrier.py`, `socket_wait.py`, `salloc_launch.sh` | node coordination + salloc helper |

scripts/vllm_dissag/README.MD

@@ -58,15 +58,16 @@ slurm gate rejects any other pairing. Dense models are TP-only; the DeepSeek fam
 
 - A Slurm cluster with `xP + yD` nodes (minimum 2: `xP=1`, `yD=1`), 8 GPUs/node.
 - A Docker image with vLLM + the KV connector(s) + NIC drivers. See **Building the Docker image**.
-- Model weights present on all allocated nodes under `/mnt/m2m_nobackup/models_blog/` or
-  `/shared_inference/models_blog/` (or set `MODEL_DIR`).
+- Model weights present on all allocated nodes at `$MODEL_DIR/$MODEL_NAME` (`MODEL_DIR` default
+  `/shared_inference/models_blog/`). The slurm resolves `MODEL_PATH` from these; override
+  `MODEL_DIR` to point elsewhere.
 - A shared filesystem for logs.
 
 ## Building the Docker image
 
-No public prebuilt image ships — build your own from the one fullsource Dockerfile (MoRI +
-AITER + vLLM + vllm-router + optional UCX/RIXL/rocSHMEM/DeepEP, from Git pins on an open
-ROCm ci_base):
+No public prebuilt image ships — build your own from the single Dockerfile
+`docker/vllm_disagg_inference.ubuntu.amd.Dockerfile` (MoRI + AITER + vLLM + vllm-router +
+optional UCX/RIXL/rocSHMEM/DeepEP, from Git pins on an open ROCm ci_base):
 
 ```bash
 cd MAD
@@ -172,10 +173,12 @@ separate port from the vLLM server.
 
 **Port defaults by connector:**
 
-| Connector | vLLM server port | Proxy port |
-|-----------|------------------|------------|
-| `rixl` (TP or DeepEP) | 2584 | `ROUTER_PORT` (30000 default) |
-| `moriio` (MoRI-EP or TP) | 20005 | `ROUTER_PORT` (30000) with the default vllm_router; 10001 with `PROXY_TYPE=moriio_toy` |
+| Connector | vLLM server port | Proxy port (`vllm_router`) | Toy-proxy port |
+|-----------|------------------|----------------------------|----------------|
+| `rixl` (TP or DeepEP) | 2584 | `ROUTER_PORT` (default 18001) | 18001 (`toy_proxy`) |
+| `moriio` (MoRI-EP or TP) | 20005 | `ROUTER_PORT` (default 30000) | 10001 (`moriio_toy`) |
+
+Override any port with the matching env var (`ROUTER_PORT`, `MORI_SERVE_PORT`, etc.).
 
 ## Proxy options
 
@@ -184,24 +187,22 @@ separate port from the vLLM server.
 
 `ROUTER_BINARY` (both connectors): path to the `vllm-router` binary. Images that bake the
 router ship it on `PATH`, so this can be left unset. Runtime-only images that do NOT ship a
-router (e.g. the mori121 MoRI-EP image) require `ROUTER_BINARY=<path>` — point it at a built
+router require `ROUTER_BINARY=<path>` — point it at a built
 `vllm-router` on a shared mount. If unset and none is on `PATH`, the launcher aborts with a
 clear error.
 
-## Validation status (image capability)
-
-The consolidated launcher assembles correct argv for all 4 combos (offline gates in `tests/`),
-but each combo also needs the connector's transport built into the image:
+## Image capability by combo
 
-| Combo | Live-validated | Image requirement |
-|-------|----------------|-------------------|
-| moriio + TP | ✅ | MoRI (mori121 image) |
-| moriio + wideEP (mori) | ✅ 1P/1D + 2P/2D | MoRI (mori121 image) |
-| rixl + TP | needs a NIXL-validated image | NixlConnector + working NCCL TP path |
-| rixl + wideEP (deepep) | needs a DeepEP-validated image | DeepEP all2all kernels |
+The launcher assembles correct argv for all 4 combos, but each combo needs its transport
+built into the image. Build with `WITH_NIXL=1` (default) to get all four; `WITH_NIXL=0` gives
+a MoRI-only image (moriio TP/wideEP).
 
-The `mori_ep_fullsource` Dockerfile builds the MoRI stack; rixl/DeepEP combos need an image
-where those transports are validated (e.g. develop's `vllm_disagg_inference` for NIXL).
+| Combo | Transport needed in image |
+|-------|---------------------------|
+| moriio + TP | MoRI |
+| moriio + wideEP (mori) | MoRI |
+| rixl + TP | NixlConnector (+ working NCCL TP path) |
+| rixl + wideEP (deepep) | DeepEP all2all kernels |
 
 ## DeepEP options (optional)
 
@@ -226,6 +227,18 @@ Parse a results log:
 python3 benchmark_parser.py <log_path>/benchmark_XXX_CONCURRENCY.log
 ```
 
+The default sweep runs `benchmark_xPyD.sh`. Two alternatives plug in via `BENCHMARK_SCRIPT_FILE`:
+
+- **`benchmark_long_context.sh`** — steady-state long-context throughput/latency with per-shape
+  warmup (each ISL/OSL cell is warmed before measurement, so JIT/cudagraph/handshake costs are
+  excluded). Use for large-ISL runs, e.g. `BENCHMARK_COMBINATIONS="28000/1000"`.
+- **`benchmark_niah.sh`** — long-context *retrieval* accuracy (see below).
+
+```bash
+export BENCHMARK_SCRIPT_FILE=benchmark_long_context.sh
+export BENCHMARK_COMBINATIONS="28000/1000" BENCHMARK_CON="1 4 8"
+```
+
 ### Long-context retrieval (NIAH)
 
 `benchmark_niah.{sh,py}` run a needle-in-a-haystack retrieval sweep (adapted from
@@ -241,20 +254,10 @@ export NIAH_WORDS="2000,8000,20000,35000"     # context sizes (words); optional
 Or run `benchmark_niah.py` standalone against any live OpenAI-compatible endpoint
 (`NIAH_URL`, `NIAH_MODEL`).
 
-## Testing the launcher (no GPUs)
-
-Offline suites validate the launcher without a cluster — run after any change to the
-launcher/connectors/yaml:
-
-```bash
-bash tests/run_all.sh             # gate_check + argv_assert; expect ALL OFFLINE SUITES PASSED
-```
-
-- `tests/gate_check.sh` — connector × WIDE_EP combo validation (valid/invalid pairings, back-compat shims).
-- `tests/argv_assert.sh` — asserts expected `vllm serve` flags/env are present/absent per cell, from
-  the launcher's `DRY_RUN=1` output (no GPUs, no container).
+## Inspecting the launcher (no GPUs)
 
-You can also dry-run any single cell to inspect the assembled command:
+Dry-run any cell to print the assembled `vllm serve` argv without launching — useful to
+sanity-check flags after editing the launcher/connectors/`models.yaml`:
 
 ```bash
 CONNECTOR=moriio WIDE_EP=1 EP_BACKEND=mori DRY_RUN=1 NODE_RANK=0 \

scripts/vllm_dissag/connectors/moriio.env

@@ -1,5 +1,5 @@
-# Platform-mandatory runtime env for the moriio (MoRIIO) connector on the
-# ROCm-7.2.3 / mori121 image. Applies to ALL moriio runs (TP and wideEP, every model).
+# Platform-mandatory runtime env for the moriio (MoRIIO) connector on ROCm 7.2.3.
+# Applies to ALL moriio runs (TP and wideEP, every model).
 #
 # These must reach the container at PID 1 — PyTorch reads the alloc-conf at import,
 # so a late shell export inside the launcher is too late. The slurm sources this file