[Bugfix] Fix accuracy issue when using flashinfer cutlass moe, TP=1 and modelopt.

[bnellnm]

Purpose

Fix accuracy issue when using flash infer cutlass moe, TP=1 and modelopt. Alternative fix to #23094.

Q: Does compressed_tensors_moe.py need a similar change?

Test Plan

VLLM_USE_FLASHINFER_MOE_FP4=1 VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=FLASHINFER lm_eval --model vllm --model_args pretrained=nvidia/Llama-4-Scout-17B-16E-Instruct-FP4,quantization=modelopt_fp4,tensor_parallel_size=1,max_model_len=2048,kv_cache_dtype=auto --gen_kwargs temperature=0.0 --limit 500 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size 200

Test w/compressed tensors

VLLM_USE_TRTLLM_ATTENTION=0 VLLM_USE_FLASHINFER_MOE_FP4=1 VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=FLASHINFER lm_eval --model vllm --model_args pretrained=RedHatAI/Llama-4-Scout-17B-16E-Instruct-NVFP4,quantization="compressed-tensors",tensor_parallel_size=1,max_model_len=2048,kv_cache_dtype=auto --gen_kwargs temperature=0.0 --limit 500 --trust_remote_code --tasks gsm8k --num_fewshot 5 --batch_size 200

Run benchmark before #22035 + after.

VLLM_USE_FLASHINFER_MOE_FP4=1 VLLM_USE_V1=1 VLLM_ATTENTION_BACKEND=FLASHINFER python3 benchmarks/benchmark_throughput.py --model=nvidia/Llama-4-Scout-17B-16E-Instruct-FP4 --quantization=modelopt_fp4 --tensor_parallel_size=1 --max_model_len=2048 --kv_cache_dtype=auto --trust_remote_code  --num-prompts 128 --input-len=1024 --output-len=1024

Test Result

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  | 0.92|±  |0.0121|
|     |       |strict-match    |     5|exact_match|↑  | 0.90|±  |0.0134|

Compressed tensors

|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.922|±  |0.0120|
|     |       |strict-match    |     5|exact_match|↑  |0.912|±  |0.0127|

Benchmark before #22035:

Throughput: 3.61 requests/s, 7386.71 total tokens/s, 3696.12 output tokens/s
Total num prompt tokens:  130876
Total num output tokens:  131072

Benchmark after:

Throughput: 3.65 requests/s, 7484.65 total tokens/s, 3742.32 output tokens/s
Total num prompt tokens:  131072
Total num output tokens:  131072

(Optional) Documentation Update

cc @varun-sundar-rabindranath , @yewentao256 , @amirkl94

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Code Review

This pull request addresses an accuracy issue with FlashInfer's CUTLASS MoE implementation when tensor parallelism is set to 1. The fix introduces a dedicated path for this configuration. However, the current implementation introduces a performance concern by repeatedly creating MoE kernels on each forward pass. My review includes suggestions to cache this kernel to avoid performance degradation, which involves refactoring the new helper function and caching the kernel within the ModelOptNvFp4FusedMoE method.

[yewentao256]

Thanks for the work! Could you also benchmark the e2e performance?

[bnellnm]

Thanks for the work! Could you also benchmark the e2e performance?

Added some benchmark results to the summary.

[bnellnm]

@mgoin, @yewentao256 , @amirkl94 do you guys know if compressed_tensors_moe.py needs a similar fix?

[mgoin]

@bnellnm yes I'm sure it needs to be updated, but I'm not sure how close it matches the current state in modelopt. It is supposed to be essentially the same, but we haven't refactored to share everything yet

[bnellnm]

@bnellnm yes I'm sure it needs to be updated, but I'm not sure how close it matches the current state in modelopt. It is supposed to be essentially the same, but we haven't refactored to share everything yet

Ok, I'll do a similar modification for compressed_tensors_moe.py. Do you know of a model I could use to verify that it's working properly?

[bnellnm]

@bnellnm yes I'm sure it needs to be updated, but I'm not sure how close it matches the current state in modelopt. It is supposed to be essentially the same, but we haven't refactored to share everything yet

Ok, I'll do a similar modification for compressed_tensors_moe.py. Do you know of a model I could use to verify that it's working properly?

Nevermind, I found one.

[mgoin]

Thank you