Add MiniCPM5 XML tool-call parser

[scrappylabsai]

Summary

Adds a tool-call parser for openbmb/MiniCPM5-1B (and its -MLX, -GGUF, -SFT variants).

The model emits tool calls in XML — not JSON — so generic parsers don't pick it up, and mlx_lm.server currently warns:

WARNING - Received tools but model does not support tool calling.

The MiniCPM5 wire format:

<function name="get_weather">
  <param name="city">Tokyo</param>
  <param name="date">2024-06-27</param>
</function>

Multi-line / special-character values may be wrapped in <![CDATA[...]]>. Attributes can be single- or double-quoted.

Implementation

• mlx_lm/tool_parsers/minicpm5.py — stateless parse_tool_call(text, tools) returning a single {"name", "arguments"} dict; raises ValueError for the ToolCallFormatter to log + drop.
• Schema-aware: rejects unknown functions, unknown params, duplicate params, missing required params, and <param> tags without a name= attribute.
• Type coercion: strings pass through; other types try strict JSON, then ast.literal_eval, then fall back to raw.
• Auto-detected by <function name= + <param name= in the chat template.

Logic ported from SGLang's MiniCPM5Detector.

Tests

tests/test_tool_parsing.py:

• Added MiniCPM5 cases to the existing multi-parser sweep (numeric args + string args).
• New test_minicpm5 covering: CDATA multi-line params, non-string typed params (array, boolean), body-only segment form, missing-required validation, unknown-param validation, duplicate-param validation, missing-name-attr validation, single-quoted attributes.

Ran 6 tests in 0.001s
OK

Smoke test

Live tested against openbmb/MiniCPM5-1B-MLX running under mlx_lm.server:

Single call:

{
  "finish_reason": "tool_calls",
  "tool_calls": [{
    "function": {
      "name": "get_weather",
      "arguments": "{\"city\": \"Tokyo\", \"date\": \"2024-06-27\"}"
    },
    "type": "function",
    "id": "..."
  }]
}

Two consecutive <function>...</function> blocks both extracted cleanly via the state machine (one segment per pair).

Test plan

• CI: existing test_tool_parsing.py continues to pass; new test_minicpm5 passes.
• Manual: mlx_lm.server --model openbmb/MiniCPM5-1B-MLX accepts tools=[...] requests without warning and returns tool_calls in the response.