Update MiniMax M3 for B300 NVFP4

[faradawn]

Mark B300 as verified for MiniMax-M3, run the NVFP4 variant on FlashInfer TRT-LLM attention with the FP8 indexer KV cache, and document the B300 agentic-coding launch with the EAGLE3-GQA draft head. Best configs based on SemiAnalysisAI/InferenceX#2328.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
vllm-recipes	Ready	Preview	Aug 5, 2026 4:25pm

[gemini-code-assist]

Code Review

This pull request adds support and verification for B300 hardware to the MiniMax-M3 model configuration, configures FlashInfer TRT-LLM attention kernels for the NVFP4 variant, and adds a guide section for agentic coding on B300. The review feedback correctly identifies that the newly added CLI arguments in extra_args use underscores (e.g., --attention_config) instead of hyphens (e.g., --attention-config), which would lead to unrecognized argument errors in vLLM.

[gemini-code-assist]

The CLI arguments should use hyphens instead of underscores to match the standard vLLM argument format (e.g., --attention-config instead of --attention_config). Using underscores will result in an unrecognized arguments error from argparse when launching the server.

Note: Please also update the other occurrences of --attention_config in this file (such as lines 170 and 227-232) to use hyphens for consistency and correctness.

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

[xinli-sw]

thanks for the changes!