Add MiniMax-M3 ROCm serving guide

[2imi9]

Adds MiniMax/MiniMax-M3.md, a serving guide for MiniMax-M3 (multimodal MoE, minimax_m3_vl, 427B total) on AMD Instinct (ROCm). There is currently only a MiniMax-M2 guide in this directory.

Verified on 8x AMD Instinct MI308X (gfx942 / CDNA3) with the MiniMax-M3-MXFP8 checkpoint on vllm/vllm-openai-rocm:v0.25.0.

It documents the ways M3 differs from the M2 series on ROCm:

• M3's sparse lightning-indexer attention is served with --attention-backend TRITON_ATTN (the M2 ROCm recipe uses ROCM_AITER_FA, which was not verified against M3's indexer here).
• Pure TP8 works (the M2 guide notes pure TP8 is unsupported).
• On gfx942 the MXFP8 checkpoint dequantizes to BF16 at load (no native MXFP8 MoE kernel), so numbers are BF16-compute.
• Long context needs --skip-mm-profiling (multimodal encoder profiling can deadlock near 128k).
• The EmbeddedLLM FP8-dynamic checkpoint needs the SwiGLU fix from vLLM #46845, else output is garbage.

Includes verified GSM8K accuracy (0.9136 flexible / 0.9121 strict, n=1319) and a baseline throughput number, with VLLM_ROCM_USE_AITER=1 / mode:3 flagged as open tuning rather than claimed as tuned results.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 22, 2026 7:35am

[gemini-code-assist]

Code Review

This pull request adds a comprehensive usage guide for serving the MiniMax-M3 multimodal model on AMD Instinct (ROCm) GPUs using vLLM. The feedback suggests improving the Docker run command by dynamically resolving the /dev/kfd group ID and adding the VLLM_USE_BREAKABLE_CUDAGRAPH=0 environment variable to optimize performance on AMD MI3xx GPUs.

[gemini-code-assist]

To make the command easier to run out-of-the-box and ensure optimal performance, we can make two improvements:

1. Replace the manual placeholder <kfd-gid> with $(stat -c '%g' /dev/kfd) to automatically resolve the group ID.
2. Add -e VLLM_USE_BREAKABLE_CUDAGRAPH=0 to prevent the decode breakable-cudagraph path from forcing eager execution on AMD MI3xx GPUs, as recommended in the model's configuration.

Suggested change

	docker run -d --name mm3-vllm \
	--network host --ipc host --shm-size 32g \
	--device /dev/kfd --device /dev/dri --group-add <kfd-gid> \
	--cap-add SYS_PTRACE --security-opt seccomp=unconfined \
	-v /path/to/MiniMax-M3-MXFP8:/models/MiniMax-M3-MXFP8:ro \
	-e VLLM_USE_V1=1 \
	--entrypoint vllm vllm/vllm-openai-rocm:v0.25.0 serve /models/MiniMax-M3-MXFP8 \
	docker run -d --name mm3-vllm \
	--network host --ipc host --shm-size 32g \
	--device /dev/kfd --device /dev/dri --group-add $(stat -c '%g' /dev/kfd) \
	--cap-add SYS_PTRACE --security-opt seccomp=unconfined \
	-v /path/to/MiniMax-M3-MXFP8:/models/MiniMax-M3-MXFP8:ro \
	-e VLLM_USE_V1=1 \
	-e VLLM_USE_BREAKABLE_CUDAGRAPH=0 \
	--entrypoint vllm vllm/vllm-openai-rocm:v0.25.0 serve /models/MiniMax-M3-MXFP8 \

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3d46b2fb7d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Move this guide into the YAML recipe

This adds a new guide under the legacy top-level Markdown tree. The repo instructions say those Markdown guides are read-only reference material and recipe updates should be made in the YAML/Next.js surface instead; keeping a separate MiniMax/MiniMax-M3.md means the legacy MkDocs page can drift from the canonical models/MiniMaxAI/MiniMax-M3.yaml command builder/API content. Please fold the ROCm-specific additions into the existing YAML guide rather than publishing another Markdown recipe.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Enable automatic tool choice with the parser

When users copy this launch command for MiniMax-M3 tool-use requests, the server is started with --tool-call-parser minimax_m3 but automatic tool calls are still disabled because --enable-auto-tool-choice is missing. vLLM's Tool Calling docs describe --enable-auto-tool-choice as mandatory for automatic function calling alongside --tool-call-parser, and the existing MiniMax-M3 YAML includes both flags, so chat requests using tool_choice=auto will not behave as advertised unless this flag is added.

Useful? React with 👍 / 👎.