feat: bf16 x mxfp4 cutlass fused moe for gpt-oss on hopper

[djmmoss]

Purpose

This PR adds the cutlass fused moe bf16xmxfp4 for gpt-oss on hopper from flashinfer the blackwell integration will come in a follow up PR soon.

The accuracy looks good, but there is a performance regression compared to triton (you can see in the results section). This is something we actively working on fixing and will push to flashinfer once it is resolved. Please let me know how you want to proceed with backend selection, and I can update accordingly.

Test Plan

Unit Tests:

python3 -m pytest tests/kernels/moe/test_mxfp4_moe.py

Accuracy Tests:

evals from here

low Ref for gpt-oss-20b and gpt-oss-120b are 0.58 and 0.68 respectively.

vllm serve gpt-oss-20b --max-model-len 32768 -tp 2
python3 -m evals --eval gpqa --model /scratch/model/gpt-oss-20b --reasoning-effort low

vllm serve gpt-oss-120b --max-model-len 32768 -tp 8
python3 -m evals --eval gpqa --model /scratch/model/gpt-oss-120b --reasoning-effort low

Performance:

vllm serve gpt-oss-120b --max-model-len 32768 -tp 8
vllm bench serve --backend vllm --model gpt-oss-120b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

vllm serve gpt-oss-20b --max-model-len 32768 -tp 2
vllm bench serve --backend vllm --model gpt-oss-20b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

Test Result

Unit Tests:

python3 -m pytest tests/kernels/moe/test_mxfp4_moe.py
================================================================================================================================= test session starts ==================================================================================================================================
platform linux -- Python 3.12.11, pytest-8.4.1, pluggy-1.6.0
rootdir: /workspace/vllm
configfile: pyproject.toml
plugins: anyio-4.10.0
collected 59 items                                                                                                                                                                                                                                                                     

tests/kernels/moe/test_mxfp4_moe.py sssssssssssssssssssssssssssssssssssssssssssssssssss........                                                                                                                                                                                  [100%]

============================================================================================================================ 8 passed, 51 skipped in 5.15s =============================================================================================================================

Accuracy Tests:

vllm serve gpt-oss-20b --max-model-len 32768 -tp 2

python3 -m evals --eval gpqa --model /scratch/model/gpt-oss-20b --reasoning-effort low
[{'eval_name': 'gpqa', 'model_name': '__scratch__model__gpt-oss-20b-low_temp1.0_20250825_121929', 'metric': 0.5732323232323232}]

vllm serve gpt-oss-120b --max-model-len 32768 -tp 8

python3 -m evals --eval gpqa --model /scratch/model/gpt-oss-120b --reasoning-effort low
[{'eval_name': 'gpqa', 'model_name': '__scratch__model__gpt-oss-120b-low_temp1.0_20250825_123502', 'metric': 0.0.6609848484848485}]

Peformance:
FlashInfer:

vllm serve gpt-oss-120b --max-model-len 32768 -tp 8
vllm bench serve --backend vllm --model gpt-oss-120b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  26.62     
Total input tokens:                      215312    
Total generated tokens:                  197144    
Request throughput (req/s):              37.56     
Output token throughput (tok/s):         7404.72   
Total Token throughput (tok/s):          15491.84  
---------------Time to First Token----------------
Mean TTFT (ms):                          3449.17   
Median TTFT (ms):                        3229.56   
P99 TTFT (ms):                           6760.19   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          94.12     
Median TPOT (ms):                        57.07     
P99 TPOT (ms):                           264.21    
---------------Inter-token Latency----------------
Mean ITL (ms):                           45.75     
Median ITL (ms):                         34.61     
P99 ITL (ms):                            220.33    
==================================================

vllm serve gpt-oss-20b --max-model-len 32768 -tp 2
vllm bench serve --backend vllm --model gpt-oss-20b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  33.94     
Total input tokens:                      215312    
Total generated tokens:                  197922    
Request throughput (req/s):              29.46     
Output token throughput (tok/s):         5831.00   
Total Token throughput (tok/s):          12174.32  
---------------Time to First Token----------------
Mean TTFT (ms):                          6173.79   
Median TTFT (ms):                        5927.93   
P99 TTFT (ms):                           11794.97  
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          149.48    
Median TPOT (ms):                        80.18     
P99 TPOT (ms):                           401.31    
---------------Inter-token Latency----------------
Mean ITL (ms):                           63.93     
Median ITL (ms):                         43.34     
P99 ITL (ms):                            401.32    
==================================================

Triton:

vllm serve gpt-oss-120b --max-model-len 32768 -tp 8
vllm bench serve --backend vllm --model gpt-oss-120b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  19.40     
Total input tokens:                      215312    
Total generated tokens:                  198101    
Request throughput (req/s):              51.54     
Output token throughput (tok/s):         10210.09  
Total Token throughput (tok/s):          21307.23  
---------------Time to First Token----------------
Mean TTFT (ms):                          4236.52   
Median TTFT (ms):                        4160.78   
P99 TTFT (ms):                           8159.43   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          81.37     
Median TPOT (ms):                        48.59     
P99 TPOT (ms):                           470.25    
---------------Inter-token Latency----------------
Mean ITL (ms):                           34.87     
Median ITL (ms):                         18.66     
P99 ITL (ms):                            76.82     
==================================================

vllm serve gpt-oss-20b --max-model-len 32768 -tp 2
vllm bench serve --backend vllm --model gpt-oss-20b --dataset-name sharegpt --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json

============ Serving Benchmark Result ============
Successful requests:                     1000      
Benchmark duration (s):                  15.66     
Total input tokens:                      215312    
Total generated tokens:                  197780    
Request throughput (req/s):              63.88     
Output token throughput (tok/s):         12633.59  
Total Token throughput (tok/s):          26387.07  
---------------Time to First Token----------------
Mean TTFT (ms):                          4105.11   
Median TTFT (ms):                        4172.38   
P99 TTFT (ms):                           5828.51   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          46.94     
Median TPOT (ms):                        29.87     
P99 TPOT (ms):                           133.46    
---------------Inter-token Latency----------------
Mean ITL (ms):                           23.34     
Median ITL (ms):                         17.03     
P99 ITL (ms):                            86.10     
==================================================

[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.

🚀

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @djmmoss.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[djmmoss]

closing as #23696 contains these changes include the blackwell changes