fix: respect tool_choice="none" by excluding tools from template

[awanawana]

Summary

Fixes #162

When tool_choice is set to "none", models like Qwen2.5-Instruct and Llama 3.x-Instruct still activate their tool-calling behavior if tools are present in the chat template context. This causes:

• finish_reason: "tool_calls" instead of "stop"
• content: null with tool_calls containing empty {} arguments
• Model encoding output data as function names instead of text content

Root Cause

The chat completions handler was passing tools to chat_kwargs without checking tool_choice. Even with tool_choice="none", if tools were present in the request, they would be passed to apply_chat_template(). Qwen2.5 and Llama 3.x have tool-calling jinja templates that activate when a tools key is present — even an empty list triggers this behavior.

Solution

Before adding tools to chat_kwargs, check if tool_choice == "none". If so, skip adding tools entirely. This prevents the template from activating tool-calling mode.

# Before
if request.tools:
    chat_kwargs["tools"] = convert_tools_for_template(request.tools)

# After
if request.tools and request.tool_choice != "none":
    chat_kwargs["tools"] = convert_tools_for_template(request.tools)

This matches the behavior of upstream vLLM's --exclude-tools-when-tool-choice-none flag.

Test Plan

# Start server
vllm-mlx serve mlx-community/Qwen2.5-14B-Instruct-4bit --port 8080

# Test with tool_choice="none"
curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "mlx-community/Qwen2.5-14B-Instruct-4bit",
    "messages": [{"role": "user", "content": "List 3 fruits as a JSON array"}],
    "tool_choice": "none",
    "tools": [],
    "max_tokens": 200
  }'

# Expected: finish_reason="stop", content contains JSON array, no tool_calls
# Before fix: finish_reason="tool_calls", content=null

🤖 Generated with Claude Code

[Thump604]

Note: our PR #173 takes a more complete approach to tool_choice=none. In addition to stripping tools from the chat template (which this PR does), #173 also suppresses tool call parsing in the response handler. Without parser suppression, the model can still generate text that looks like a tool call and the parser will extract it as one, even though the user requested tool_choice=none.

The two fixes are complementary but independent: template stripping prevents the model from seeing tool definitions, parser suppression prevents false-positive tool call extraction from generated text.