vllm main branch (upstream/main commit 0008729ab)

---

vllm serve <model> --reasoning-parser glm45 --enable-auto-tool-choice --tool-call-parser hermes

---

import openai
client = openai.OpenAI(base_url="http://localhost:8000/v1")

response = client.chat.completions.create(
    model="<model>",
    messages=[{"role": "user", "content": "What's the weather in Beijing?"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "get_weather",
            "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}
        }
    }],
    tool_choice={"type": "function", "function": {"name": "get_weather"}}
)

---

AssertionError
  File "vllm/entrypoints/openai/engine/serving.py", line 615, in _parse_tool_calls_from_content
    assert content is not None

---

# Line 615 — ToolChoiceFunction branch
if request.tool_choice and isinstance(request.tool_choice, ToolChoiceFunction):
    assert content is not None  # <-- CRASH when reasoning_parser consumed all content

# Line 624 — ChatCompletionNamedToolChoiceParam branch  
elif request.tool_choice and isinstance(request.tool_choice, ChatCompletionNamedToolChoiceParam):
    assert content is not None  # <-- CRASH

---

elif (request.tool_choice and type(request.tool_choice) is ChatCompletionNamedToolChoiceParam):
    assert tool_calls is not None and len(tool_calls) > 0  # <-- CRASH (downstream)

---

if request.tool_choice and isinstance(request.tool_choice, ToolChoiceFunction):
    if not content:
        logger.warning(
            "Forced tool_choice=%s but content is empty "
            "(reasoning_parser may have consumed all output). "
            "Skipping tool call parsing.",
            request.tool_choice.name,
        )
        return None, content
    # ... original logic

---

elif type(request.tool_choice) is ChatCompletionNamedToolChoiceParam:
    if not tool_calls:
        message = ChatMessage(role=role, reasoning=reasoning, content=content or "")
    else:
        # ... original tool_call construction logic