Update MiniMax M3 MXFP4 AITER command

[haic0]

Refresh the MI355X MXFP4 guide command to use the ROCm AITER MoE path and text-only launch flags.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 3, 2026 3:17am

[gemini-code-assist]

Code Review

This pull request updates the vllm serve guide command for the MiniMax-M3 model to use ROCm AITER MoE paths, text-only launch flags, and associated environment variables, while removing the multimodal encoder arguments. The review feedback highlights two main issues: first, the $MAX_MODEL_LEN environment variable is used in the command but not defined in the snippet; second, the mxfp4 variant hardware overrides still contain the old multimodal encoder arguments and lack the new environment variables and arguments, which should be updated to maintain consistency.

[gemini-code-assist]

The command uses the $MAX_MODEL_LEN environment variable on line 469, but it is not defined or initialized anywhere in this bash snippet. If a user copy-pastes and runs this block directly, the variable will evaluate to an empty string, causing the vllm serve command to fail with a parsing error.

Please define MAX_MODEL_LEN in the environment setup lines above the command.

  export VLLM_ROCM_USE_AITER=1
  export VLLM_ROCM_USE_AITER_MOE=1
  export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1
  export MAX_MODEL_LEN=131072

[gemini-code-assist]

The updated guide command now uses the text-only launch flags (--language-model-only, --no-enable-prefix-caching) and the ROCm AITER MoE path (--moe-backend aiter), and removes the multimodal encoder flags (--mm-encoder-tp-mode data, --mm-encoder-attn-backend ROCM_AITER_FA).

However, the mxfp4 variant definition under variants.mxfp4.hardware_overrides.mi355x (lines 123-132) still contains the old multimodal encoder arguments and lacks the new environment variables and arguments:

    hardware_overrides:
      mi355x:
        docker_image: "vllm/vllm-openai-rocm:nightly"
        extra_args:
          - "--trust-remote-code"
          # Keep the multimodal encoder loaded and run it data-parallel on MI355X.
          - "--mm-encoder-tp-mode"
          - "data"
          - "--mm-encoder-attn-backend"
          - "ROCM_AITER_FA"

To ensure that the automated command builder/UI generates the correct command for the mxfp4 variant, please update the variant definition in lines 123-132 to match the new guide command:

    hardware_overrides:
      mi355x:
        docker_image: "vllm/vllm-openai-rocm:nightly"
        extra_args:
          - "--trust-remote-code"
          - "--no-enable-prefix-caching"
          - "--language-model-only"
          - "--moe-backend"
          - "aiter"
        extra_env:
          VLLM_ROCM_USE_AITER: "1"
          VLLM_ROCM_USE_AITER_MOE: "1"
          VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS: "1"