[None][fix] Fix moe backend setting#15861
Conversation
Signed-off-by: shuyixiong <219646547+shuyixiong@users.noreply.github.com>
59fff7c to
a2b6e66
Compare
📝 WalkthroughWalkthroughRemoved a fallback conditional in ChangesMoE Backend Selection
Estimated code review effort: 2 (Simple) | ~10 minutes 🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
|
/bot run --disable-fail-fast |
|
PR_Github #57096 [ run ] triggered by Bot. Commit: |
|
PR_Github #57096 [ run ] completed with state
|
|
/bot run --disable-fail-fast |
1 similar comment
|
/bot run --disable-fail-fast |
|
PR_Github #57202 [ run ] triggered by Bot. Commit: |
|
PR_Github #57202 [ run ] completed with state
|
|
/bot run --disable-fail-fast |
|
PR_Github #57383 [ run ] triggered by Bot. Commit: |
|
PR_Github #57383 [ run ] completed with state
|
Summary by CodeRabbit
resolve_moe_clsunconditionally downgraded an unquantized (BF16) MoE fromTRTLLMGenFusedMoEtoCutlassFusedMoE, somoe_config.backend="TRTLLM"was silently ignored for BF16 checkpoints. This PR removes that downgrade so the user-requested TRTLLM backend is actually used on its supported BF16 path.The downgrade rule (
if moe_cls == TRTLLMGenFusedMoE and not has_quant: moe_cls = CutlassFusedMoE) was originally added as a fallback: some BF16 cases lacked FlashInfer support, so it fell back to Cutlass for those. It was gated by_supports_flashinfer_bf16_routing_method(routing_method), which keptTRTLLMGenFusedMoEwhenever the FlashInfer BF16 path was actually available. That guard was dropped in [TRTLLM-13052][feat] Enable TRTLLM moe backend for nemotron-h BF16 ckpt #14944, turning a conditional fallback into an unconditional downgrade that discards a valid TRTLLM selection.Description
Test Coverage
PR Checklist
Please review the following before submitting your PR:
PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.
PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.
Test cases are provided for new code paths (see test instructions)
If PR introduces API changes, an appropriate PR label is added - either
api-compatibleorapi-breaking. Forapi-breaking, includeBREAKINGin the PR title.Any new dependencies have been scanned for license and vulnerabilities
CODEOWNERS updated if ownership changes
Documentation updated as needed
Update tava architecture diagram if there is a significant design change in PR.
The reviewers assigned automatically/manually are appropriate for the PR.
Please check this after reviewing the above items as appropriate for this PR.
GitHub Bot Help
To see a list of available CI bot commands, please comment
/bot help.