Add B200 GPU configuration for MiniMax-M2.7

[kjiang249]

No description provided.

[gemini-code-assist]

Code Review

This pull request updates the documentation to include support for NVIDIA B200 GPUs and provides a recommended configuration for running the MiniMax-M2.7 model on them. Feedback suggests including the --compilation-config flag to enable specific optimizations and adjusting the reasoning-parser value to maintain consistency with other examples in the guide.

[gemini-code-assist]

The B200 configuration is missing the --compilation-config which includes the fuse_minimax_qk_norm optimization recommended for this model series (as noted in line 184). Additionally, the reasoning-parser value minimax_m2_append_think is inconsistent with all other examples in this guide (e.g., lines 70, 126, 153) which use minimax_m2. It is recommended to maintain consistency across the documentation unless this specific parser is required for B200.

Suggested change

	VLLM_FLOAT32_MATMUL_PRECISION="high" vllm serve MiniMaxAI/MiniMax-M2.7 \
	--trust-remote-code \
	--tensor-parallel-size 4 \
	--enable-auto-tool-choice \
	--tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2_append_think
	VLLM_FLOAT32_MATMUL_PRECISION="high" vllm serve MiniMaxAI/MiniMax-M2.7 \
	--tensor-parallel-size 4 \
	--tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2 \
	--compilation-config '{"mode":3,"pass_config":{"fuse_minimax_qk_norm":true}}' \
	--enable-auto-tool-choice \
	--trust-remote-code

[simon-mo]

This is quite different from the previous configuration

vllm serve MiniMaxAI/MiniMax-M2.7 \
  --tensor-parallel-size 4 \
  --tool-call-parser minimax_m2 \
  --reasoning-parser minimax_m2  \
  --compilation-config '{"mode":3,"pass_config":{"fuse_minimax_qk_norm":true}}' \
  --enable-auto-tool-choice \
  --trust-remote-code

Have you tested that,

• we need minimax_m2_append_think
• we cannot do --compilation-config '{"mode":3,"pass_config":{"fuse_minimax_qk_norm":true}}' on blackwell?

[wzhao18]

We can remove minimax_m2_append_think and add --compilation-config '{"mode":3,"pass_config":{"fuse_minimax_qk_norm":true}}'. Though the fuse_minimax_qk_norm is not yet available on latest vllm release.

[simon-mo]

My main concern is we have different configurations for Hpper and Blackwell and I would like to educate the users on the minimal difference between the two and explain the rationale.

[wzhao18]

yes we will keep them consistent. Thanks for flagging this out.

[simon-mo]

Thanks. please ping me on Slack when it's updated, thx