[https://nvbugs/6095421][fix] In ModelConfig.resolve_moe_backend, after the GptOss branch, return "DEEPGEMM" w

[tensorrt-cicd]

Summary

• Root cause: AUTO moe_backend resolves to CUTLASS on Blackwell, whose FP8_BLOCK_SCALES path uses a DeepGEMM JIT kernel that only supports Hopper, so the DeepSeek-V3-Lite-fp8 worker throws during warmup and the disagg parent deadlocks waiting for MPI_READY.
• Fix: In ModelConfig.resolve_moe_backend, after the GptOss branch, return "DEEPGEMM" when is_sm_100f(sm_version) is True so non-GptOss MoE models use the Blackwell-native FP8 backend.
• 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/6095421

Summary by CodeRabbit

• New Features
  • Added optimized MoE backend support for Blackwell GPU architecture to improve inference performance.

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR adds Blackwell GPU-specific MOE backend selection to TensorRT-LLM's model configuration. It imports a GPU detection utility and adds a conditional branch in resolve_moe_backend that returns "DEEPGEMM" when targeting Blackwell SM 100F devices with the GptOssForCausalLM architecture and AUTO backend mode.

Changes

Blackwell MOE Backend Selection

Layer / File(s)	Summary
Blackwell GPU MOE backend selection tensorrt_llm/_torch/model_config.py	Import is_sm_100f GPU detection utility and add conditional logic in ModelConfig.resolve_moe_backend to return "DEEPGEMM" when is_sm_100f() is true for GptOssForCausalLM models under AUTO backend selection.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~8 minutes

Suggested reviewers

• StanleySun639
• hyukn

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the specific change: returning DEEPGEMM in ModelConfig.resolve_moe_backend for Blackwell (SM100f), directly matching the main code modification.
Description check	✅ Passed	The PR description provides root cause, fix explanation, test coverage verification, and bug link, but lacks formal PR checklist completion and detailed API impact assessment.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.
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

Warning

Review ran into problems

🔥 Problems

Git: Failed to clone repository. Please run the @coderabbitai full review command to re-trigger a full review. If the issue persists, set path_filters to include or exclude specific files.

---

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

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

tensorrt_llm/_torch/model_config.py (1)

1-1: ⚠️ Potential issue | 🟠 Major | ⚡ Quick win

Add/update the required NVIDIA copyright header on this modified Python source file.

This file appears to be missing the required TensorRT-LLM copyright/license header for modified source files.

As per coding guidelines "**/*.{cpp,cc,cxx,h,hpp,hxx,cu,py}: All TensorRT-LLM source files must contain an NVIDIA copyright header with year of latest modification, using the Apache 2.0 license text."

🤖 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/model_config.py` at line 1, This file
tensorrt_llm/_torch/model_config.py is missing the required NVIDIA Apache-2.0
copyright header; add the standard TensorRT-LLM Apache 2.0 license header (with
the year of latest modification) at the very top of the file before any imports
(i.e., above the existing import contextlib line) so the file contains the
official multi-line copyright and license block used across TensorRT-LLM Python
sources.

🤖 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.

Outside diff comments:
In `@tensorrt_llm/_torch/model_config.py`:
- Line 1: This file tensorrt_llm/_torch/model_config.py is missing the required
NVIDIA Apache-2.0 copyright header; add the standard TensorRT-LLM Apache 2.0
license header (with the year of latest modification) at the very top of the
file before any imports (i.e., above the existing import contextlib line) so the
file contains the official multi-line copyright and license block used across
TensorRT-LLM Python sources.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: ec34a9d1-93e2-4de0-8bfa-5f8af6107ddc

📥 Commits

Reviewing files that changed from the base of the PR and between 8a11da8 and b311270.

📒 Files selected for processing (1)

• tensorrt_llm/_torch/model_config.py

[heyuhhh]

/bot run

[tensorrt-cicd]

PR_Github #48841 [ run ] triggered by Bot. Commit: 721c243 Link to invocation

[tensorrt-cicd]

PR_Github #48841 [ run ] completed with state SUCCESS. Commit: 721c243
/LLM/main/L0_MergeRequest_PR pipeline #38598 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[heyuhhh]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #48859 [ run ] triggered by Bot. Commit: 721c243 Link to invocation

[tensorrt-cicd]

PR_Github #48859 [ run ] completed with state ABORTED. Commit: 721c243

Link to invocation