feat: Add MiniMax-M2 tool call parser

[janhilgard]

Summary

• Add MiniMaxToolParser for MiniMax-M2 models' native XML tool call format
• Integrate tool parser with reasoning parser path in streaming mode
• Strip MiniMax special tokens from output

Details

MiniMax-M2 models use a custom XML-based tool call format distinct from other parsers. This PR adds:

1. minimax_tool_parser.py — Full parser with regex-based extraction of <minimax:tool_call> blocks, <invoke> elements, and <parameter> elements. Supports both non-streaming (extract_tool_calls) and streaming (extract_tool_calls_streaming) modes. Parameter values are auto-parsed as JSON when possible for proper typing.

2. CLI integration — minimax added to --tool-call-parser choices.

3. Special token stripping — MiniMax end-of-turn token [e~[, role markers ]~b]assistant etc., and BOS token ]~!b[ are added to the global SPECIAL_TOKENS_PATTERN regex.

4. Streaming: reasoning + tool parser integration — MiniMax wraps tool calls inside <think> blocks, so when --reasoning-parser is enabled, the reasoning parser captures tool call XML as reasoning content. This PR adds detection of tool call markers in the reasoning stream and redirects them to the content stream for proper tool parser processing. This fixes streaming tool calls when both --reasoning-parser and --tool-call-parser minimax are used together.

Changes

File	Change
vllm_mlx/tool_parsers/minimax_tool_parser.py	New parser for MiniMax XML format
vllm_mlx/tool_parsers/__init__.py	Import and register MiniMaxToolParser
vllm_mlx/cli.py	Add "minimax" to --tool-call-parser choices
vllm_mlx/api/utils.py	Add MiniMax special tokens to SPECIAL_TOKENS_PATTERN
vllm_mlx/server.py	Integrate tool parser within reasoning parser streaming path

Test plan

• Tested with MiniMax-M2-5-REAP-39 (8-bit, 138GB MoE) on Apple M3 Ultra
• Non-streaming tool calls parse correctly
• Streaming tool calls: no XML leaking, proper tool_calls delta emission
• Streaming with --reasoning-parser deepseek_r1: tool calls detected in reasoning and redirected
• Streaming content (no tool call): reasoning/content split works correctly
• Special tokens stripped from responses

Usage

vllm-mlx serve <minimax-model> \
    --enable-auto-tool-choice \
    --tool-call-parser minimax \
    --reasoning-parser deepseek_r1

🤖 Generated with Claude Code

[raullenchai]

Thanks for this PR. This works out really well on M3 ultra with "MiniMax-M2.5-MLX-4bit".

python -m vllm_mlx.cli serve "~\MiniMax-M2.5-MLX-4bit" \
  --host 0.0.0.0 \
  --port 8000 \
  --enable-prefix-cache \
  --max-num-seqs 2 \
  --tool-call-parser minimax \
  --enable-auto-tool-choice