Add MiniMax-M2.5 B200 FP4/NVFP4 serving recipe

[faradawn]

Summary

• Add ### B200 (FP4 / NVFP4) section to MiniMax/MiniMax-M2.5.md
• Base config: TP=2, FP8 KV cache, --max-cudagraph-capture-size 2048, --stream-interval 20, --no-enable-prefix-caching
• Expert parallelism variant: TP=4 + --enable-expert-parallel for higher concurrency (EP degree matches TP per benchmark sweep)

Based on SemiAnalysisAI/InferenceX#996

Test plan

• Verify nvidia/MiniMax-M2.5-NVFP4 model loads with TP=2 on B200
• Verify expert-parallel variant with TP=4 + --enable-expert-parallel

[gemini-code-assist]

Code Review

This pull request adds a new section to the MiniMax-M2.5 documentation for B200 GPUs, providing command examples for the FP4 quantized model. The review feedback correctly identifies that the vllm serve commands are missing essential tool-calling and reasoning parser flags required for the model's specialized features. Additionally, the non-standard --no-enable-prefix-caching flag should be removed to ensure the commands are valid and prevent startup errors.

[gemini-code-assist]

The vllm serve command is missing the tool-calling and reasoning parser flags (--tool-call-parser, --reasoning-parser, and --enable-auto-tool-choice) which are included in other MiniMax-M2.5 examples in this file. Without these, the model's specialized capabilities will not function correctly. Additionally, --no-enable-prefix-caching is not a standard vLLM flag; prefix caching is disabled by default, so this flag should be removed to avoid potential startup errors.

Suggested change

	vllm serve nvidia/MiniMax-M2.5-NVFP4 \
	--tensor-parallel-size 2 \
	--gpu-memory-utilization 0.90 \
	--kv-cache-dtype fp8 \
	--max-cudagraph-capture-size 2048 \
	--stream-interval 20 \
	--no-enable-prefix-caching \
	--trust-remote-code
	vllm serve nvidia/MiniMax-M2.5-NVFP4 \
	--tensor-parallel-size 2 \
	--gpu-memory-utilization 0.90 \
	--kv-cache-dtype fp8 \
	--max-cudagraph-capture-size 2048 \
	--stream-interval 20 \
	--tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2_append_think \
	--enable-auto-tool-choice \
	--trust-remote-code

[gemini-code-assist]

This example is also missing the required parsers for tool-calling and reasoning, and includes the non-standard --no-enable-prefix-caching flag. Including the parsers ensures the model behaves consistently with the other configurations provided in this guide.

Suggested change

	vllm serve nvidia/MiniMax-M2.5-NVFP4 \
	--tensor-parallel-size 4 \
	--enable-expert-parallel \
	--gpu-memory-utilization 0.90 \
	--kv-cache-dtype fp8 \
	--max-cudagraph-capture-size 2048 \
	--stream-interval 20 \
	--no-enable-prefix-caching \
	--trust-remote-code
	vllm serve nvidia/MiniMax-M2.5-NVFP4 \
	--tensor-parallel-size 4 \
	--enable-expert-parallel \
	--gpu-memory-utilization 0.90 \
	--kv-cache-dtype fp8 \
	--max-cudagraph-capture-size 2048 \
	--stream-interval 20 \
	--tool-call-parser minimax_m2 \
	--reasoning-parser minimax_m2_append_think \
	--enable-auto-tool-choice \
	--trust-remote-code