[LoRA] Set default MXFP4 LoRA backend to Marlin

[xyang16]

Purpose

This PR set default MXFP4 LoRA backend to Marlin because Triton has accuracy issues and Marlin has slight better performance.

• Use Triton only if Marlin is disabled (set VLLM_MXFP4_USE_MARLIN=0 explicitly) and triton_kernels is supported.
• Use Marlin by default
  • if VLLM_MXFP4_USE_MARLIN is not set
  • if VLLM_MXFP4_USE_MARLIN=1
  • if triton_kernels is not supported

Benchmarking

Marlin:

VLLM_MXFP4_USE_MARLIN=1 vllm serve openai/gpt-oss-20b \
  --tensor-parallel-size 1 \
  --max-num-seqs 16 \
  --compilation-config '{"cudagraph_mode": "PIECEWISE", "compile_sizes": [1, 2, 4, 8, 16]}' \
  --enable-lora \
  --max-loras 1 \
  --lora-modules lora1=/opt/dlami/nvme/models/gpt-oss-20b-lora-gsm8k \
  --max-lora-rank 32 \
  --no-enable-prefix-caching

vllm bench serve \
  --model openai/gpt-oss-20b \
  --lora-modules lora1 \
  --dataset-name sharegpt \
  --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json \
  --sharegpt-output-len 800 \
  --max-concurrency 16 \
  --num-prompts 1000 \
  --num-warmups 60 \
  --ignore-eos

============ Serving Benchmark Result ============
Successful requests:                     1000      
Failed requests:                         0         
Maximum request concurrency:             16        
Benchmark duration (s):                  418.75    
Total input tokens:                      226792    
Total generated tokens:                  800000    
Request throughput (req/s):              2.39      
Output token throughput (tok/s):         1910.47   
Peak output token throughput (tok/s):    2053.00   
Peak concurrent requests:                32.00     
Total token throughput (tok/s):          2452.06   
---------------Time to First Token----------------
Mean TTFT (ms):                          70.91     
Median TTFT (ms):                        60.91     
P99 TTFT (ms):                           204.49    
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          8.24      
Median TPOT (ms):                        8.25      
P99 TPOT (ms):                           8.50      
---------------Inter-token Latency----------------
Mean ITL (ms):                           8.24      
Median ITL (ms):                         8.11      
P99 ITL (ms):                            9.11      
==================================================

Triton:

vllm serve openai/gpt-oss-20b \
  --tensor-parallel-size 1 \
  --max-num-seqs 16 \
  --compilation-config '{"cudagraph_mode": "PIECEWISE", "compile_sizes": [1, 2, 4, 8, 16]}' \
  --enable-lora \
  --max-loras 1 \
  --lora-modules lora1=/opt/dlami/nvme/models/gpt-oss-20b-lora-gsm8k \
  --max-lora-rank 32 \
  --no-enable-prefix-caching

vllm bench serve \
  --model openai/gpt-oss-20b \
  --lora-modules lora1 \
  --dataset-name sharegpt \
  --dataset-path ShareGPT_V3_unfiltered_cleaned_split.json \
  --sharegpt-output-len 800 \
  --max-concurrency 16 \
  --num-prompts 1000 \
  --num-warmups 60 \
  --ignore-eos

============ Serving Benchmark Result ============
Successful requests:                     1000      
Failed requests:                         0         
Maximum request concurrency:             16        
Benchmark duration (s):                  439.55    
Total input tokens:                      226792    
Total generated tokens:                  800000    
Request throughput (req/s):              2.28      
Output token throughput (tok/s):         1820.06   
Peak output token throughput (tok/s):    1968.00   
Peak concurrent requests:                32.00     
Total token throughput (tok/s):          2336.02   
---------------Time to First Token----------------
Mean TTFT (ms):                          75.77     
Median TTFT (ms):                        49.08     
P99 TTFT (ms):                           214.07    
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          8.65      
Median TPOT (ms):                        8.66      
P99 TPOT (ms):                           8.94      
---------------Inter-token Latency----------------
Mean ITL (ms):                           8.65      
Median ITL (ms):                         8.50      
P99 ITL (ms):                            9.49      
==================================================

Marlin is slightly better because in mxfp4 Triton LoRA is implemented in UnfusedOAITritonExperts. It has to unfuse the activation and reduction to allow to inject lora modules, so this makes it lose the triton_kernels's optimizations of fused activation and fused moe_sum.

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request changes the default MXFP4 backend for LoRA from Triton to Marlin, citing better performance and accuracy. The logic is updated to select Marlin by default, and only fall back to Triton if VLLM_MXFP4_USE_MARLIN is explicitly set to 0 and Triton kernels are supported. The change is correct, well-scoped to LoRA as per the PR title, and aligns with the stated purpose. The implementation is clear and concise.

[yewentao256]

Please add metrics in the PR description to show the acc / perf issue you mentioned.

1. lm_eval for acc
2. vllm bench... for performance

[jeejeelee]

After addressing @yewentao256's comment, LGTM

[xyang16]

Please add metrics in the PR description to show the acc / perf issue you mentioned.

1. lm_eval for acc
2. vllm bench... for performance

@yewentao256 Thanks for review!

1. Accuracy: Currently gpt-oss mxfp4 with LoRA + triton kernel is generating garbage output, see [Bug]: FULL_AND_PIECEWISE cudagraph mode leading to !!! in generated text #29539 (comment)
2. Performance: I have pasted vllm bench numbers in the description.

[yewentao256]

Please add metrics in the PR description to show the acc / perf issue you mentioned.

1. lm_eval for acc
2. vllm bench... for performance

@yewentao256 Thanks for review!

1. Accuracy: Currently gpt-oss mxfp4 with LoRA + triton kernel is generating garbage output, see [Bug]: FULL_AND_PIECEWISE cudagraph mode leading to !!! in generated text #29539 (comment)
2. Performance: I have pasted vllm bench numbers in the description.

Let's fix the issue first then, do you have time to take a deep look into this issue?

[xyang16]

Let's fix the issue first then, do you have time to take a deep look into this issue?

I raised a PR to fix this: #30585

There are PRs to fix the cudagraph issue: #30650, #30887