vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
     --dtype bfloat16 --max-model-len 4096 --enforce-eager \
     --limit-mm-per-prompt '{"image": 1}'

---

def test_tokenize_then_chat_completion_with_image(
    server: RemoteOpenAIServer,
    local_asset_server,
):
    """Tokenize a multimodal message, then send the same message to chat
    completions.  The chat completion must succeed (not 500)."""

    image_url = local_asset_server.url_for("stop_sign.jpg")
    messages = [
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": image_url}},
                {"type": "text", "text": "Describe this image briefly."},
            ],
        }
    ]

    # Step 1: tokenize (this triggers multimodal processing in the renderer)
    tok_resp = requests.post(
        server.url_for("tokenize"),
        json={"model": MODEL_NAME, "messages": messages},
    )
    tok_resp.raise_for_status()
    tok_data = tok_resp.json()
    assert tok_data["count"] > 0, "Tokenization must return tokens"

    # Step 2: chat completion with the SAME multimodal message
    chat_resp = requests.post(
        server.url_for("v1/chat/completions"),
        json={
            "model": MODEL_NAME,
            "messages": messages,
            "max_tokens": 10,
            "temperature": 0.0,
        },
    )

    assert chat_resp.status_code == 200, (
        f"Chat completion failed after tokenize: "
        f"status={chat_resp.status_code}, body={chat_resp.text}"
    )
    chat_data = chat_resp.json()
    assert chat_data["choices"][0]["message"]["content"], (
        "Chat completion must produce non-empty content"
    )