MiniMax-M2.5: update B200 FP8 serving config

[faradawn]

Summary

• Add --enable-expert-parallel (tp:4/ep:4 validated; tp:2/ep:2 also noted)
• Add --gpu-memory-utilization 0.90
• Add --block-size 32
• Add --kv-cache-dtype fp8
• Add --stream-interval 20
• Add --no-enable-prefix-caching
• Switch image tag from nightly to latest

Based on SemiAnalysisAI/InferenceX#1010, which validated and widened the B200 FP8 search space for MiniMax-M2.5 (tp:4/ep:4 at conc 256–512, tp:2/ep:2 at conc 512).

[gemini-code-assist]

Code Review

This pull request updates the MiniMax-M2.5 documentation for B200 GPU configurations, transitioning to the latest vLLM image and adding several performance-tuning flags. However, the proposed configuration contains a parallelism inconsistency where setting tensor parallelism to 4 on a 4-GPU setup makes expert parallelism redundant. Furthermore, the update inadvertently removed critical tool-calling and reasoning parser flags and introduced an invalid argument, '--no-enable-prefix-caching', which would cause server startup failures.

[gemini-code-assist]

There is an inconsistency between the text, the command, and the PR description regarding parallelism:

• The text states the recommended configuration uses 4 GPUs.
• The command below uses --tensor-parallel-size 4.
• In vLLM, if the number of GPUs equals the tensor parallel size, expert parallelism is effectively disabled (EP=1), making --enable-expert-parallel redundant.
• The PR description mentions tp:4/ep:4 was validated, which requires 16 GPUs.

If the intention is to provide a 4-GPU recommendation that utilizes expert parallelism, the configuration should likely be --tensor-parallel-size 2 (which defaults to EP=2 on 4 GPUs). Please clarify the intended hardware target.

[gemini-code-assist]

This update introduces two issues:

1. Regression: The tool-calling and reasoning parser flags (--tool-call-parser, --reasoning-parser, and --enable-auto-tool-choice) have been removed. These are essential for the model's specialized features, such as structured tool use and the <think> block formatting.
2. Invalid Argument: --no-enable-prefix-caching is not a valid vLLM argument. Prefix caching is disabled by default in vLLM; including an unrecognized flag will cause the server to fail at startup. If you wish to ensure it is off, simply omit the --enable-prefix-caching flag.

Suggested change

	--enable-expert-parallel \
	--gpu-memory-utilization 0.90 \
	--block-size 32 \
	--kv-cache-dtype fp8 \
	--stream-interval 20 \
	--no-enable-prefix-caching \
	--trust-remote-code
	--enable-expert-parallel \
	--gpu-memory-utilization 0.90 \
	--block-size 32 \
	--kv-cache-dtype fp8 \
	--stream-interval 20 \
	--tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2_append_think \
	--enable-auto-tool-choice \
	--trust-remote-code