Repository navigation
[quantization] Preserve explicit CPU embedding domains - #1108
Merged
Merged
Conversation
Keep row-wise UINT8 embedding DQ/GATHER/Q and scaled activation boundaries explicit in Circle. Prepare selected-row reference CPU plans without restoring the full table, preserving shared integer weights and the output contract. Add native round-trip, arithmetic, memory, sharing and regression coverage, plus a reproducible CPU benchmark and documentation of runtime limitations. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Record the successful full CPU suite at implementation commit 934c059: 4,849 tests run, 4,691 passed, 158 skipped and exit code 0. Replace the prior incomplete-validation statement while retaining the history of interrupted attempts and the existing target-runtime limitations. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
…ard propagation Keep the calibrated arguments and metadata of an explicitly preserved CPU embedding Quantize. Let downstream Concat legalization insert a separate conversion rather than erase the original rounding boundary by retargeting it. Cover direct and layout paths with native Circle file round trips and retain unprotected private-quantizer retargeting. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
…nge clamps Finish reduced-range clamp insertion before classifying DQ consumers. Include new and previously owned range clamps in the CPU island so traversal order and repeat execution cannot remove the DQ feeding a float clamp. Test the real export, legalization, Circle file round trip and native execution for consecutive boundaries, saturation, unclipped values and MUL. Exercise an adverse iteration order without replacing production passes or kernels. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Normalize reduced-range test imports and record the current full CPU run: 4,858 tests, 4,700 passed, 158 skipped, no failures/errors and exit code 0. Keep the pre-fix counterexamples and production-runtime limitations explicit. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Replace implementation history and experiment reports with the rationale for explicit Circle conversions, selected-row execution and preserved numerical boundaries. Describe storage/ABI invariants and supported runtime contracts, and update the referring documentation labels. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Keep explicit Q/DQ boundaries through float insertion layouts and independent constant square transforms. Respect float tensors during qparam propagation, and add explicit conversions when those tensors enter integer concat domains. Reject unsupported mixed side branches and keep shared-table projections under the existing FC policy. Add real tiny Llama and Qwen3-VL export/file-reload/native integration tests, including nonidentity SpinQuant rotation, multimodal insertion, frozen-grid oracles, selected-row allocation checks and explicit decoder stage handoffs. Generalize the CPU embedding design documentation and its support matrix. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Replay layout-only paths on the original activation codes for ordinary quantized FC input edges, keeping shared float consumers and the FC's weight and output domains intact. Require existing weight/output quantization so genuine float projections retain their native float contract. Complete shared float side-input layout closure before protecting terminal quantizers, preventing their Q/DQ boundaries from folding into float layouts. Add real Circle save/reload, native/fake arithmetic, shared-consumer, range, creation-order and pass-lifecycle regressions, and document the boundaries. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Recognize per-tensor output quantization through reshape and permute before restoring an ordinary quantized FC's input layouts on the original A codes. Follow only layout data edges and retain the existing weight-domain gate; stop at arithmetic, concat, DQ and unsupported output layouts. Extend real Circle file round-trip coverage with a nonidentity output permute, layout chains, shared float consumers, native input-code extraction, reduced-range clipping and independent fake-quant arithmetic. Document the supported output-contract search without changing pass order or input grids. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
Add preserve_embedding_quantization=False to CompileConfigV1 and select the existing preservation pass only for an explicit True, before quant folding. Forward the compile policy through recipe exporters and the export CLI. Cover legacy and opt-in serialized contracts across public conversion APIs, YAML and real recipe exports; retain explicit embedding numerical regressions. Document the opt-in policy and its separation from runtime execution settings. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
mhs4670go
force-pushed
the
cpu-explicit-embedding
branch
from
November 5, 2026 08:01
9f2e595 to
b6c3c1d
Compare
Document the embedding Quantize retarget exception in the system design. Separate the canonical lookup graph from optional scaled post-processing, and distinguish stage-conversion semantics from the DQ/Q test implementation. TICO-DCO-1.0-Signed-off-by: seongwoo <mhs4670go@naver.com>
mhs4670go
marked this pull request as ready for review
November 6, 2026 02:46
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Keep row-wise UINT8 embedding DQ/GATHER/Q and scaled activation boundaries
explicit in Circle. Prepare selected-row reference CPU plans without restoring
the full table, preserving shared integer weights and the output contract.
Add native round-trip, arithmetic, memory, sharing and regression coverage,
plus a reproducible CPU benchmark and documentation of runtime limitations.
TICO-DCO-1.0-Signed-off-by: seongwoo mhs4670go@naver.com