vllm - 💡(How to fix) Fix [Bug]: Responses API does not surface reasoning output with `--reasoning-parser gemma4` (works with deepseek_r1) [1 pull requests]

OS                           : Ubuntu 22.04.5 LTS (x86_64)
PyTorch version              : 2.11.0+cu130
CUDA used to build PyTorch   : 13.0
Python version               : 3.12.13
Is CUDA available            : True
CUDA runtime version         : 13.0.88
GPU 0-3: NVIDIA RTX PRO 6000 Blackwell Server Edition
vLLM Version                 : 0.21.0
transformers                 : 5.8.1

---

docker run -d \
  --name vllm-gemma4 \
  --gpus all \
  --ipc=host \
  -p 8000:8000 \
  -e VLLM_ENABLE_RESPONSES_API_STORE=1 \
  vllm/vllm-openai:v0.21.0 \
  --model google/gemma-4-26B-A4B-it \
  --tensor-parallel-size 4 \
  --max-model-len 32768 \
  --gpu-memory-utilization 0.9 \
  --max-num-seqs 4 \
  --max-num-batched-tokens 14336 \
  --tool-call-parser functiongemma \
  --enable-auto-tool-choice \
  --reasoning-parser gemma4

---

import requests, json

BASE = "http://localhost:8000/v1"
MODEL = "google/gemma-4-26B-A4B-it"

tools = [
    {
        "type": "function",
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    }
]

# Test 1: Chat Completions — reasoning WORKS
print("=== Chat Completions (works) ===")
r1 = requests.post(f"{BASE}/chat/completions", json={
    "model": MODEL,
    "messages": [{"role": "user", "content": "What is the weather in NYC?"}],
    "tools": [{"type": "function", "function": {"name": "get_weather", "description": "Get weather", "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}}}],
    "tool_choice": "required",
    "max_tokens": 1024,
    "chat_template_kwargs": {"enable_thinking": True},
})
d1 = r1.json()
msg = d1["choices"][0]["message"]
print(f"  reasoning: {msg.get('reasoning', '')[:200]}")
print(f"  tool_calls: {msg.get('tool_calls')}")
print(f"  finish_reason: {d1['choices'][0]['finish_reason']}")

# Test 2: Responses API — reasoning NOT surfaced
print("\n=== Responses API (broken) ===")
r2 = requests.post(f"{BASE}/responses", json={
    "model": MODEL,
    "input": [{"role": "user", "content": "What is the weather in NYC?"}],
    "tools": tools,
    "tool_choice": "required",
    "reasoning": {"effort": "high"},
})
d2 = r2.json()
usage = d2.get("usage", {})
output_details = usage.get("output_tokens_details", {})
print(f"  reasoning_tokens: {output_details.get('reasoning_tokens', 0)}")
print(f"  output types: {[item.get('type') for item in d2.get('output', [])]}")
print(f"  top-level reasoning field: {d2.get('reasoning')}")

# Test 3: Responses API text-only — still no reasoning
print("\n=== Responses API text-only (also broken) ===")
r3 = requests.post(f"{BASE}/responses", json={
    "model": MODEL,
    "input": [{"role": "user", "content": "What is 2+2? Think step by step."}],
    "reasoning": {"effort": "high"},
    "max_output_tokens": 500,
})
d3 = r3.json()
usage3 = d3.get("usage", {})
print(f"  reasoning_tokens: {usage3.get('output_tokens_details', {}).get('reasoning_tokens', 0)}")
print(f"  output types: {[item.get('type') for item in d3.get('output', [])]}")

---

=== Chat Completions (works) ===
  reasoning: The user is asking about the weather in NYC. I should look at the available tools... The `get_weather` tool seems appropriate for this task.
  tool_calls: [{'id': 'chatcmpl-tool-...', 'type': 'function', 'function': {'name': 'get_weather', 'arguments': '{"city": "NYC"}'}}]
  finish_reason: tool_calls

=== Responses API (broken) ===
  reasoning_tokens: 0
  output types: ['function_call']
  top-level reasoning field: None

=== Responses API text-only (also broken) ===
  reasoning_tokens: 0
  output types: ['message']

---

=== Chat Completions with reasoning ===
  reasoning present: True
  reasoning preview: The user wants to know the product of 15 × 37...
  finish_reason: stop

=== Chat Completions with tools + reasoning ===
  reasoning present: True
  reasoning preview: The user is asking about the weather in NYC. I should look for a tool...
  tool_calls: [{"function": {"name": "get_weather", "arguments": "{\"city\": \"NYC\"}"}}]
  finish_reason: tool_calls

=== Responses API with reasoning={'effort': 'high'} ===
  reasoning_tokens: 0
  output types: ['message']
  has ResponseReasoningItem: False

=== Responses API with tools + reasoning ===
  reasoning_tokens: 0
  output types: ['function_call']
  has ResponseReasoningItem: False