MiniMax-M3 spec decoding Blackwell

[faradawn]

Use Eagle3-GQA draft head and enable FLASHINFER + TRT-LLM attention for spec_decoding on Blackwell (B200/B300/GB200/GB300). Non-Blackwell NVIDIA stays on the standard Eagle3 head; AMD stays on TRITON_ATTN. Based on SemiAnalysisAI/InferenceX#2328 and SemiAnalysisAI/InferenceX#2337.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview, Comment	Jul 28, 2026 1:54pm

[gemini-code-assist]

Code Review

This pull request introduces hardware overrides for the 'blackwell' architecture in the MiniMax-M3 model configuration, specifying speculative decoding and attention settings. The review feedback correctly identifies that the CLI arguments --attention_config.backend and --attention_config.use_trtllm_attention are non-standard in vLLM and may cause server startup failures, recommending standard alternatives.

[gemini-code-assist]

The CLI arguments --attention_config.backend and --attention_config.use_trtllm_attention are not standard vLLM arguments and will likely cause the server to fail with an unrecognized arguments error.

1. --attention_config.backend should be --attention-backend.
2. --attention_config.use_trtllm_attention is not a standard vLLM CLI argument. Please verify if this is a custom argument supported by your specific Docker image (vllm/vllm-openai:minimax-m3) or if it should be passed as an environment variable instead.

          - "--attention-backend"
          - "FLASHINFER"
          - "--attention_config.use_trtllm_attention"
          - "true"