mi355 fp4 minimax vllm single node

[functionstackx]

@claude vllm add minimax fp4 for mi355 tp8

huggingface model amd/MiniMax-M2.1-MXFP4

amd doesnt have minimax m2.5 mxfp4 yet, just do minimax m2.1 mxfp4 for now to test it in WIP PR since it is same arch https://huggingface.co/amd/MiniMax-M2.1-MXFP4

[Klaud-Cold]

Claude finished @functionstackx's task in 4m 12s —— View job • claude/issue-826-20260301-0402 • Create PR ➔

---

Adding MiniMax M2.1 MXFP4 for MI355 (vLLM)

• Gather context (read CLAUDE.md, existing configs)
• Understand existing MiniMax/AMD config patterns
• Create benchmark script minimaxm2.1_fp4_mi355x.sh
• Add minimaxm2.1-fp4-mi355x-vllm config to amd-master.yaml
• Update perf-changelog.yaml
• Commit and push changes
• Provide PR link

Summary

Added MiniMax M2.1 MXFP4 benchmark configuration for MI355x with vLLM:

Config: minimaxm2.1-fp4-mi355x-vllm

• Model: amd/MiniMax-M2.1-MXFP4
• Image: vllm/vllm-openai-rocm:v0.16.0
• TP: 8
• Seq lengths: 1k1k, 1k8k, 8k1k (conc 4-64)

Benchmark script: benchmarks/single_node/minimaxm2.1_fp4_mi355x.sh

• Based on existing MiniMax FP8 script with --trust-remote-code and --block-size=32
• Added FP4-relevant AITer env vars (VLLM_ROCM_USE_AITER=1, VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1, VLLM_ROCM_USE_AITER_MHA=0)

|

[functionstackx]

@claude no tp=8, do tp=2,4

also only do export VLLM_ROCM_USE_AITER=1 & not VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION=1, VLLM_ROCM_USE_AITER_MHA=0 similar to the andy luo recipe for minimax fp8

plz refer to benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh

[Klaud-Cold]

Claude finished @functionstackx's task in 3m 17s —— View job • claude/issue-826-20260301-0409

---

Updating MiniMax M2.1 MXFP4 for MI355 (vLLM)

• Read reference script minimaxm2.5_fp8_mi355x.sh and existing configs
• Create benchmark script with correct env vars (only VLLM_ROCM_USE_AITER=1)
• Add config to amd-master.yaml with TP=2,4
• Update perf-changelog.yaml
• Commit and push changes

Summary

Added MiniMax M2.1 MXFP4 benchmark for MI355x vLLM, matching the minimaxm2.5_fp8_mi355x pattern:

Config: minimaxm2.1-fp4-mi355x-vllm

• Model: amd/MiniMax-M2.1-MXFP4
• Image: vllm/vllm-openai-rocm:v0.16.0
• TP: 2, 4 (not 8)
• Env: Only VLLM_ROCM_USE_AITER=1 (no UNIFIED_ATTENTION or MHA vars)
• Seq lengths: 1k1k, 1k8k, 8k1k (conc 4-64)

Benchmark script: benchmarks/single_node/minimaxm2.1_fp4_mi355x.sh — identical to minimaxm2.5_fp8_mi355x.sh

Create PR
| claude/issue-826-20260301-0409