[Hotfix] Fix router gemm on sm103

[Fridge003]

Motivation

Modifications

Accuracy Tests

Speed Tests and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review and Merge Process

1. Ping Merge Oncalls to start the process. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

[Fridge003]

/rerun-test test_deepseek_v3_fp4_4gpu.py

[github-actions]

✅ 4-gpu-b200 (1 test): View workflow run

cd test/ && python3 registered/quant/test_deepseek_v3_fp4_4gpu.py

[gemini-code-assist]

Code Review

This pull request modifies the DeepSeek-V2 model implementation to restrict a specialized FlashInfer router GEMM kernel optimization specifically to SM 100 devices. The reviewer noted that this change might unnecessarily exclude other Blackwell variants from the optimization and suggested either using a more inclusive check or adding documentation to justify the restriction.

[gemini-code-assist]

The change from _device_sm >= 100 to _device_sm == 100 effectively disables the specialized FlashInfer router GEMM kernel for Blackwell variants other than SM 100 (such as SM 103 mentioned in the PR title). While this correctly addresses the reported issue on SM 103 by falling back to the dsv3_router_gemm kernel, it also excludes any other potential Blackwell variants (e.g., SM 101) from this optimization. If other variants are known to be compatible, consider using a more inclusive check or adding a comment explaining why this optimization is strictly limited to SM 100.