[Bugfix] Add SM110/SM120 device capability checks for NVFP4 MoE backends

[Code4me2]

Summary

Extend device capability checks to include SM110 and SM120 GPU families in NVFP4 MoE backend selection code that was missed in PR #33417.

Problem

After PR #33417, NVFP4 MoE models still fail on RTX Blackwell GPUs (SM120) and DGX Spark (SM121) with errors like:

RuntimeError: FlashInfer-CUTLASS MoE kernel does not support current device

Root Cause

PR #33417 updated flashinfer_cutlass_moe.py and cutlass_moe.py but missed these files which still only check for SM100:

• flashinfer_fp4_moe.py
• flashinfer_trtllm_moe.py
• flashinfer_cutedsl_moe.py
• flashinfer_utils.py

Solution

Add explicit family checks for SM100/110/120, matching the approach in the files updated by #33417:

p.is_device_capability_family(100)
or p.is_device_capability_family(110)
or p.is_device_capability_family(120)

This enables support for:

• SM100-109: Blackwell data center (B100, B200)
• SM110-119: Future Blackwell variants
• SM120-129: Blackwell consumer/workstation (RTX 5090, DGX Spark GB10)

Files Changed

• vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py
• vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py
• vllm/model_executor/layers/fused_moe/flashinfer_cutedsl_moe.py
• vllm/model_executor/layers/quantization/utils/flashinfer_utils.py

Testing

Tested on RTX 5090 (SM120) and DGX Spark GB10 (SM121) with nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4

Related Issues

Fixes #31085
Fixes #30135
Fixes #29141
Related to #28589

[dosubot]

Related Documentation

No published documentation to review for changes on this repository.

Write your first living document

^{How did I do? Any feedback?}  

[gemini-code-assist]

Code Review

This pull request correctly extends device capability checks to include SM110 and SM120 GPU families, fixing the reported bug. The changes are applied across all necessary files. However, the logic for checking the device capability is duplicated in multiple places. This code duplication is a maintainability risk, as it was the root cause of the issue this PR is fixing (some files were missed in a previous update). I've added suggestions to refactor this duplicated logic using any() for better readability and to make future updates easier and less error-prone.

[gemini-code-assist]

This device capability check is duplicated across multiple files. To improve maintainability and avoid future bugs where one location is updated but others are missed, consider using any() with a generator expression. This makes the code more concise and easier to update.

        return p.is_cuda() and any(p.is_device_capability_family(sm) for sm in (100, 110, 120))

[gemini-code-assist]

This device capability check is duplicated across multiple files. To improve maintainability and avoid future bugs where one location is updated but others are missed, consider using any() with a generator expression. This makes the code more concise and easier to update.

    return p.is_cuda() and any(p.is_device_capability_family(sm) for sm in (100, 110, 120))

[gemini-code-assist]

This device capability check is duplicated across multiple files. To improve maintainability and avoid future bugs where one location is updated but others are missed, consider using any() with a generator expression. This makes the code more concise and easier to update.

    return p.is_cuda() and any(p.is_device_capability_family(sm) for sm in (100, 110, 120))

[gemini-code-assist]

This device capability check is duplicated across multiple files. To improve maintainability and avoid future bugs where one location is updated but others are missed, consider using any() with a generator expression. This makes the code more concise and easier to update.

    is_blackwell = any(current_platform.is_device_capability_family(sm) for sm in (100, 110, 120))

[johnnynunez]

thanks! LGTM @mgoin

[mgoin]

@Code4me2 how is this correct? It was intentional to avoid sm120 for flashinfer_trtllm_moe and flashinfer_cutedsl_moe as these are unsupported on sm120. Maybe sm110 can be included but I need more specific validation.

Regarding this problem:

After PR #33417, NVFP4 MoE models still fail on RTX Blackwell GPUs (SM120) and DGX Spark (SM121) with errors like:

RuntimeError: FlashInfer-CUTLASS MoE kernel does not support current device

I'm not sure how these changes resolve that since they don't affect flashinfer_cutlass_moe at all. Can you give more context?

[Code4me2]

@mgoin yeah sorry about this one. The series of the 3 PRs were me trying to get nemotron nano NVFP4 working on the dgx spark and RTX 5090, and I tried multiple different things. This was a lapse in judgement, because it was the other two PRs that fixed the issue. Apologies. Closing now.

[mgoin]

No worries I thought I was missing something, thanks!