Fix LFM2.5 tool parser inference

[blairhudson]

LFM2.5 tokenizers expose <|tool_call_start|> and <|tool_call_end|> tokens, but do not provide chat-template metadata that lets mlx-lm infer a tool parser. As a result, server responses can emit raw tool-call markup in message.content instead of OpenAI-compatible tool_calls.

This adds tokenizer-vocab based parser inference for that marker pair and maps it to the existing pythonic tool parser.

Tests:

• uv run --with-editable . --with unittest-xml-reporting python -m unittest tests.test_tool_parsing tests.test_tokenizers.TestTokenizers.test_tool_calling

You can try it yourself with:

uv tool install "git+https://github.com/blairhudson/mlx-lm.git@fix/lfm25-tool-parser-inference"
mlx_lm.server --model LiquidAI/LFM2.5-1.2B-Instruct-MLX-8bit

[jbuchananr]

Can you add support for this use case:

Reproduce

  mlx_lm.server --model <your-mlx-model> --port 8080

  curl -s http://localhost:8080/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
      "model": "<your-mlx-model>",
      "temperature": 0.0,
      "max_tokens": 2048,
      "tool_choice": "auto",
      "messages": [
        {
          "role": "user",
          "content": "Order the ingredients for a lasagna to be delivered to 845 Willow Lane, 
  Springfield, IL 62704. Include noodles, ground beef, ricotta, mozzarella, parmesan, tomato sauce, 
  onion, garlic, olive oil, basil, oregano, and salt."
        }
      ],
      "tools": [
        {
          "type": "function",
          "function": {
            "name": "grocery.orderIngredients",
            "description": "Orders a list of ingredients for delivery to a specified address.",
            "parameters": {
              "type": "object",
              "properties": {
                "ingredientList": {
                  "type": "array",
                  "description": "List of ingredients to order.",
                  "items": {
                    "type": "object",
                    "properties": {
                      "name":   { "type": "string" },
                      "amount": { "type": "number" },
                      "unit":   { "type": "string" }
                    },
                    "required": ["name", "amount", "unit"]
                  }
                },
                "deliveryAddress": { "type": "string" }
              },
              "required": ["ingredientList", "deliveryAddress"],
              "additionalProperties": false
            }
          }
        }
      ]
    }'

  Expected response (OpenAI-shaped)

  {
    "choices": [{
      "finish_reason": "tool_calls",
      "message": {
        "role": "assistant",
        "content": null,
        "tool_calls": [{
          "id": "...",
          "type": "function",
          "function": {
            "name": "grocery.orderIngredients",
            "arguments": "{\"ingredientList\":[{\"name\":\"noodles\",\"amount\":500,\"unit\":\"g\"}, 
  ...],\"deliveryAddress\":\"845 Willow Lane, Springfield, IL 62704\"}"
          }
        }]
      }
    }]
  }

  Actual response

  {
    "choices": [{
      "finish_reason": "tool_calls",
      "message": {
        "role": "assistant",
        "content": "I am placing an order for the following ingredients to be delivered to 845 Willow 
  Lane, Springfield, IL 62704: 500g noodles, 300g ground beef, 200g ricotta, 250g mozzarella, 100g 
  parmesan, 400ml tomato sauce, 100g onion, 50g garlic, 30ml olive oil, 10g basil, 5g oregano, 5g 
  salt."
      }
    }],
    "usage": { "prompt_tokens": 256, "completion_tokens": 482, "total_tokens": 738 }
  }

  Why this looks like a server bug, not a model regression

  - finish_reason is "tool_calls" — the server detected a tool call happened.
  }

  Actual response

  {
    "choices": [{
      "finish_reason": "tool_calls",
      "message": {
        "role": "assistant",
        "content": "I am placing an order for the following ingredients to be delivered to 845 Willow
  Lane, Springfield, IL 62704: 500g noodles, 300g ground beef, 200g ricotta, 250g mozzarella, 100g
  parmesan, 400ml tomato sauce, 100g onion, 50g garlic, 30ml olive oil, 10g basil, 5g oregano, 5g
  salt."
      }
    }],
    "usage": { "prompt_tokens": 256, "completion_tokens": 482, "total_tokens": 738 }
  }

Issue

• finish_reason is "tool_calls" — the server detected a tool call happened.
• But message.tool_calls is missing, and the natural-language form of the arguments leaked into
  message.content.
• A simpler tool call (single string arg) on the same model/server returns a correctly structured
  tool_calls array — only the nested-array schema trips it.

This is the signature of a tool-call parser that handles flat arguments but doesn't extract the
model's native <tool_call>{...}</tool_call> block when the JSON payload contains a nested array of
objects.

Env

Environment

• mlx-lm 0.31.3
• macOS 15.5, Apple Silicon
• temperature: 0.0, tool_choice: "auto"
• LiquidAI/LFM2.5-1.2B-Instruct-MLX-8bit