Minimax M3 disagg

[faradawn]

Set backend FLASHINFER, use_trtllm_attention, and indexer_kv_dtype fp8 in the attention config for MiniMax-M3 pd_cluster (disaggregated P/D) serving on Blackwell (B200/B300/GB200/GB300). Based on SemiAnalysisAI/InferenceX#2230, SemiAnalysisAI/InferenceX#2310, SemiAnalysisAI/InferenceX#2338, SemiAnalysisAI/InferenceX#2339, and SemiAnalysisAI/InferenceX#2340.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 27, 2026 10:54pm

[gemini-code-assist]

Code Review

This pull request adds strategy overrides for the pd_cluster on blackwell hardware in models/MiniMaxAI/MiniMax-M3.yaml. A review comment correctly identifies that the command-line arguments in extra_args use underscores instead of hyphens, which will cause unrecognized argument errors in vLLM, and provides a suggestion to fix them.

[gemini-code-assist]

Command-line arguments in vLLM use hyphens (-) rather than underscores (_) for parameter names. Using underscores like --attention_config.backend will result in an unrecognized argument error when launching the server. They should be updated to use hyphens, matching the standard CLI convention used throughout the rest of this configuration file.

          - "--attention-config.backend"
          - "FLASHINFER"
          - "--attention-config.use-trtllm-attention"
          - "true"
          - "--attention-config.indexer-kv-dtype"
          - "fp8"