[https://nvbugs/6212252][fix] Lift the SM list onto DeepGemmFusedMoE._SUPPORTED_SM_VERSIONS = {100, 103} and

[tensorrt-cicd]

Summary

• Root cause: get_moe_cls() unconditionally returns DeepGemmFusedMoE for backend='DEEPGEMM'; on H20 (SM90) the Blackwell-only DeepGEMM kernel runs anyway and asserts sfa_dtype == kFloat.
• Fix: Lift the SM list onto DeepGemmFusedMoE._SUPPORTED_SM_VERSIONS = {100, 103} and add an SM check in create_moe.py's DEEPGEMM branch with fallback to CutlassFusedMoE (which supports FP8_BLOCK_SCALES on SM90), mirroring DENSEGEMM/CUTEDSL.
• Automated fix generated by repair-bot

Test plan

• Verify fix on the same GPU type as the original failure
• Check for regressions in related tests

Links

• Bug: https://nvbugs/6212252

Summary by CodeRabbit

• Bug Fixes
  • Enhanced Mixture of Experts (MoE) backend compatibility by validating GPU architecture versions during initialization.
  • Implemented automatic fallback to compatible alternatives when the optimized backend is unsupported on detected hardware.
  • Added warning notifications to inform users when hardware is incompatible with specialized implementations.

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR centralizes SM version compatibility checking for the DeepGEMM MoE backend. It extracts the supported SM versions into a class constant and uses that constant both in backend-selection fallback logic and in the existing can_implement validation method.

Changes

DeepGEMM SM Compatibility Gating

Layer / File(s)	Summary
SM support constant and validation tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py	DeepGemmFusedMoE._SUPPORTED_SM_VERSIONS is declared as {100, 103}. The can_implement method is refactored to check the detected SM version against this constant and emit an updated warning message.
Backend selection with SM fallback tensorrt_llm/_torch/modules/fused_moe/create_moe.py	get_moe_cls now queries the current SM version when selecting the DEEPGEMM backend; if unsupported, it logs a warning and falls back to CutlassFusedMoE.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Possibly related PRs

• NVIDIA/TensorRT-LLM#14362: Both PRs update DeepGEMM Python-side SM compatibility gating to treat SM103 as supported in DeepGEMM-based MoE backend selection and validation.

Suggested reviewers

• xxi-nv
• yuxianq
• mikeiovine

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly describes the main change: lifting the SM list onto a class constant and adding SM compatibility checking in the DEEPGEMM backend, matching the actual code modifications.
Description check	✅ Passed	The PR description includes root cause, solution details, test plan with checkmarks, and bug link, satisfying the template requirements despite missing some optional PR checklist items.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 0

🧹 Nitpick comments (1)

tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py (1)

380-380: ⚡ Quick win

Use frozenset instead of set for immutable class constant.

_SUPPORTED_SM_VERSIONS is named as a constant but uses a mutable set. Replace with frozenset to enforce immutability and align with the constant naming convention.

🔒 Proposed fix

-    _SUPPORTED_SM_VERSIONS = {100, 103}
+    _SUPPORTED_SM_VERSIONS = frozenset({100, 103})

As per coding guidelines, Python constants should be immutable. The static analysis tool flagged this as RUF012 (mutable default value for class attribute).

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py` at line 380,
Replace the mutable set assigned to the class constant _SUPPORTED_SM_VERSIONS
with an immutable frozenset to enforce immutability and follow constant naming
conventions; locate the declaration of _SUPPORTED_SM_VERSIONS in
fused_moe_deepgemm.py and change the RHS from {100, 103} to use frozenset({100,
103}) (or frozenset((100, 103))) so the constant cannot be mutated at runtime.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py`:
- Line 380: Replace the mutable set assigned to the class constant
_SUPPORTED_SM_VERSIONS with an immutable frozenset to enforce immutability and
follow constant naming conventions; locate the declaration of
_SUPPORTED_SM_VERSIONS in fused_moe_deepgemm.py and change the RHS from {100,
103} to use frozenset({100, 103}) (or frozenset((100, 103))) so the constant
cannot be mutated at runtime.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 04f18277-6509-4e14-ac9a-914fdb23b535

📥 Commits

Reviewing files that changed from the base of the PR and between da9db3a and 5ab45fe.

📒 Files selected for processing (2)

• tensorrt_llm/_torch/modules/fused_moe/create_moe.py
• tensorrt_llm/_torch/modules/fused_moe/fused_moe_deepgemm.py

[xxi-nv]

Please add another constraint: only FP8_BLOCK QUANT on HOPPER is supported on CUTLASS.

[xxi-nv]

Or a suggested way to fix is to specify the CUTLASS MOE backend in the test case instead of the fallback.