[Bugfix][SM70] Default audited MTP MoE tiles on - #391
Merged
yangzhuxinyzx merged 1 commit intoAug 28, 2026
Conversation
Signed-off-by: yangzhuxinyzx <153831768+yangzhuxinyzx@users.noreply.github.com>
yangzhuxinyzx
marked this pull request as ready for review
August 28, 2026 07:43
yangzhuxinyzx
merged commit Aug 28, 2026
2db84d5
into
codex/v100-qwen38-mtp4-public-20260828-1455
4 checks passed
yangzhuxinyzx
deleted the
agent/v100-repair-pr389-default-on-20260828-072543
branch
August 28, 2026 07:49
This was referenced Aug 28, 2026
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Purpose\n\nRepair parent PR #389 before integration.\n\n- Enable the audited exact-shape SM70 MTP MoE tiles by default.\n- Keep VLLM_SM70_MTP_MOE_TUNED_CONFIG=0 as an explicit rollback.\n- Keep dispatch generic and shape-bounded; no model or checkpoint identity gate is added.\n- Align the benchmark policy, environment tests, and design records with the runtime default.\n\n## Base\n\n- Parent PR head: ae560f5\n- Repair head: ff2602c\n- Integration target after parent merge: main at 03c04da\n\n## Test Plan\n\nRun the focused environment, benchmark-policy, MoE-config, Qwen3.8 weight-loading, MTP, and MRV2 warmup tests, then run Ruff and whitespace checks.\n\n## Test Result\n\n- Focused tests: 126 passed, 2 benign SWIG deprecation warnings.\n- Ruff check: pass.\n- Ruff format check: pass.\n- git diff --check: pass.\n- Parent GPU evidence is unchanged because this repair changes only the default gate: Qwen3.8 M5 5.26x, M1 1.56x, bitwise exact; full-model HumanEval 8/8 and 138.26 tok/s.\n- The existing Qwen3.6 M1 rejection remains excluded by the exact M2-M16 shape gate; unmatched shapes retain the 0.0.3 fallback.\n\n## Risk and rollback\n\nThe runtime only selects the two audited SM70 TP4 local-shape families. Set VLLM_SM70_MTP_MOE_TUNED_CONFIG=0 to restore the previous default immediately.