Minimax M2.5 MXFP4 benchmark for MI355x vLLM v0.19.1 (TP=1,2,4)

[functionstackx]

Add MiniMax M2.5 MXFP4 benchmark config for MI355x with vLLM v0.17.1, now that AMD's MXFP4 checkpoint is out: https://huggingface.co/amd/MiniMax-M2.5-MXFP4

• Model: amd/MiniMax-M2.5-MXFP4
• Image: vllm/vllm-openai-rocm:v0.17.1
• TP=2 and TP=4 (matching MiniMax M2.5 FP8 pattern)
• VLLM_ROCM_USE_AITER=1, with AITER MoE fallback for TP>=4
• Seq lengths: 1k1k, 1k8k, 8k1k (conc 4-64)

Closes #826

Generated with Claude Code

[functionstackx]

@claude can u comment out TP=4 and only do TP=2 for now since TP=4 is bugged vllm-project/vllm#35637

[Klaud-Cold]

Claude finished @functionstackx's task in 53s —— View job

---

Commented out TP=4 for MiniMax M2.1 MXFP4

Due to vllm-project/vllm#35637, TP=4 has been disabled. Only TP=2 will run.

• Read current benchmark files
• Comment out TP=4 in .github/configs/amd-master.yaml (all 3 seq-len configs)
• Update perf-changelog.yaml with PR link and description
• Commit and push changes

Changes made:

• .github/configs/amd-master.yaml: Commented out all 3 tp: 4 search-space entries under minimaxm2.1-fp4-mi355x-vllm with a reference to the vLLM issue
• perf-changelog.yaml: Updated entry to reflect TP=2 only and added PR Minimax M2.5 MXFP4 benchmark for MI355x vLLM v0.19.1 (TP=1,2,4) #827 link
  |

[functionstackx]

@chunfangamd i am seeing good mxfp4 performance when comparing to FP8. just waiting for AMD to release mxfp4 version of MiniMax M2.5 now. currently AMD only have mxfp4 version of MiniMax M2.1.

[chunfangamd]

Looks good to me

[seungrokj]

@benenzhu vllm/vllm-openai-rocm:v0.19.1 is ready, can you submit a PR ? cc. @ajith-sirra-amd @functionstackx

[functionstackx]

Feel free to ping me for quick review when u have a PR for minimax mxfp4 that is working and has passed pr validation and there is an vllm-projects/recipes update (if needed)

[benenzhu]

https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24647578553/job/72063351980?pr=827

@seungrokj Hi, I retriggerd the CI, seems the mia1-p01-g12 machine's huggingface dir is broken now. Can't load the model. Only this machine will fail

[functionstackx]

@chunfangamd can u take a look at debugging this? Probably previous runs have corrupted the huggingface model cache from @benenzhu testing. The fix is just rm this minimax ckpt folder on the cluster

[benenzhu]

@chunfangamd can u take a look at debugging this? Probably previous runs have corrupted the huggingface model cache from @benenzhu testing. The fix is just rm this minimax ckpt folder on the cluster

@functionstackx Hi I think that's some gh runner problem in this machine, the docker command also fails.

[functionstackx]

@chunfangamd can u take a look at debugging this? Probably previous runs have corrupted the huggingface model cache from @benenzhu testing. The fix is just rm this minimax ckpt folder on the cluster

@functionstackx Hi I think that's some gh runner problem in this machine, the docker command also fails.

Both @chunfangamd and/or @cquil11 have cluster access to fix it. @chunfangamd can u help?

[benenzhu]

@functionstackx Hi, I have disable the bad machine and the actions have passed. May you help review it?

1. Same script as minimax fp8 and no additional flags. So no need update to vllm recipes.
2. Accuracy result:

https://inferencex.semianalysis.com/inference?unofficialRun=24648550630&g_rundate=2026-04-17&g_runid=24588340987&g_model=MiniMax-M2.5

[functionstackx]

@cquil11 or @Oseltamivir csn u review this pr

[functionstackx]

I am the original author of this pull request so GitHub doesn't let me approve my own pr

[functionstackx]

@benenzhu one issue I see before I go to sleep is that vllm-project/recipes doesn't have an recipe for this minimax M2.5 mxfp4. We wanna ensure that the entire ml community benefits from ur hard work. Can u please create an recipe for it in that documentation

[benenzhu]

@benenzhu one issue I see before I go to sleep is that vllm-project/recipes doesn't have an recipe for this minimax M2.5 mxfp4. We wanna ensure that the entire ml community benefits from ur hard work. Can u please create an recipe for it in that documentation

Yeah, thanks for the review. I will raise one PR for this.

FP4 uses a model path of amd/MiniMax-M2.5-MXFP4, so should let ppl know that our model's path.

Have a good night.

[benenzhu]

vllm-project/recipes#300

[functionstackx]

Thanks @benenzhu ! Can u please merge this Kimi K2.5 mxfp4 recipe? It's been stuck unmerged for 3 weeks.

vllm-project/recipes#296

[functionstackx]

@chunfangamd it seems to fail
https://github.com/SemiAnalysisAI/InferenceX/actions/runs/24661496941/job/72108418033?pr=827