[Bug] Qwen2.5-VL-72B image input not working in SGLang, works fine in vLLM

[qWaitCrypto]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.
• 4. If the issue you raised is not a bug but a question, please raise a discussion at https://github.com/sgl-project/sglang/discussions/new/choose Otherwise, it will be closed.
• 5. Please use English, otherwise it will be closed.

Describe the bug

🧾 Description:

When deploying qwen2.5-vl-72b-awq using SGLang, image inputs (via image_url) are not correctly handled. The same prompt works as expected in vLLM, where the model successfully describes the image.

Reproduction

✅ Reproduction Steps:

✅ SGLang Launch Command:

python -m sglang.launch_server \
  --model-path qwen-vl-72b \
  --port 30000 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --mem-fraction-static 0.8 \
  --tp 4 \
  --tool-call-parser qwen25

✅ OpenAI-Compatible API Call (cURL):

curl -X POST "http://0.0.0.0:30000/v1/chat/completions" \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen2.5-vl",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "describe this picture"
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://dashscope.oss-cn-beijing.aliyuncs.com/images/dog_and_girl.jpeg"
            }
          }
        ]
      }
    ],
    "top_p": 0.8
  }'

---

🧾 SGLang Response:

{
  "id": "803d3c01743b4429b61c0a83d60eda5b",
  "object": "chat.completion",
  "created": 1742528000,
  "model": "qwen2.5-vl",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "I'm sorry, but I cannot see any picture attached to your message. Could you please provide more information or upload the picture again? I'll do my best to describe it for you."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 22,
    "completion_tokens": 41,
    "total_tokens": 63
  }
}

---

✅ Comparison with vLLM:

Using the exact same model and cURL request, the image is successfully described in vLLM deployment. This confirms that the issue is not with the prompt or model, but with how SGLang handles image_url type content in the message payload.

---

📌 Expected Behavior:

SGLang should support OpenAI-compatible image inputs by correctly parsing messages.content[].image_url.url and feeding the image into the model’s visual encoder.

Environment

🧪 Environment:

• Model: qwen2.5-vl-72b-awq
• Deployment: SGLang 0.4.4.post1
• API Protocol: OpenAI-compatible Chat Completions API
• vLLM Behavior: ✅ Working as expected

[shuaills]

@woodx9 Can you have a look of this, thanks

[woodx9]

@shuaills Sure, plz assign it to me.

[qWaitCrypto]

0.0

[m0g1cian]

+1.
We've also noticed significant behavior difference between SGLang and vLLM when running Qwen2.5 VL 72B instruct (unquantized). When we sent multiple images to the SGLang backend, the model sometimes replies "I cannot see any picture attached to your message" too, which is not observed on vLLM backend or the offical Qwen2.5 VL 72B web demo on huggingface.

[Titan-p]

+1. We've also noticed significant behavior difference between SGLang and vLLM when running Qwen2.5 VL 72B instruct (unquantized). When we sent multiple images to the SGLang backend, the model sometimes replies "I cannot see any picture attached to your message" too, which is not observed on vLLM backend or the offical Qwen2.5 VL 72B web demo on huggingface.

Two possible causes

1. Related to [Bug] Inaccurate or Inconsistent Output in Qwen2.5-VL Multi-Image Testing with sglang #4123 fix: second_per_grid_ts should be used to get mrope position #3682. This should be fixed after this pr merged.
2. qwen vl currently do not work with radix cache . So --disable-radix-cache & --chunked-prefill-size -1 should work.

[adarshxs]

Please run the server with the flag --chat-template qwen2-vl

Refer: https://docs.sglang.ai/backend/openai_api_vision.html#Chat-Template

If you do not specify a vision model’s --chat-template, the server uses Hugging Face’s default template, which only supports text.

[woodx9]

works fine for me on Qwen2.5-VL-7B. @qWaitCrypto plz give it a try.

:~/workspace# curl -X POST "http://0.0.0.0:30020/v1/chat/completions"   -H "Content-Type: application/json"   -d '{
    "model": "Qwen/Qwen2.5-VL-7B-Instruct",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "describe this picture"
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://dashscope.oss-cn-beijing.aliyuncs.com/images/dog_and_girl.jpeg"
            }
          }
        ]
      }
    ],
    "top_p": 0.8
  }'
{"id":"b95acb7bb4cc48f9a83c90d6f3a58ebd","object":"chat.completion","created":1742632799,"model":"Qwen/Qwen2.5-VL-7B-Instruct","choices":[{"index":0,"message":{"role":"assistant","content":"The image depicts a heartwarming scene of a woman and her dog enjoying a moment together on a sandy beach. The woman, dressed in a plaid shirt and jeans, is sitting on the sand and extending her hand for a high-five with her dog. The dog, a light-colored breed, is sitting upright and reaching up with its paw to meet the woman's hand. The setting is serene, with the ocean in the background and the sun setting, casting a warm, golden glow over the scene. The overall atmosphere is one of joy, companionship, and the special bond between the woman and her pet.","reasoning_content":null,"tool_calls":null},"logprobs":null,"finish_reason":"stop","matched_stop":151645}],"usage":{"prompt_tokens":3601,"total_tokens":3726,"completion_tokens":125,"prompt_tokens_details":null}

[qWaitCrypto]

When I set --disable-radix-cache & --chunked-prefill-size -1 & --chat-template qwen2-vl, I can read images normally
thx

[M0gician]

When I set --disable-radix-cache & --chunked-prefill-size -1 & --chat-template qwen2-vl, I can read images normally
thx

Could you try multiple images? We found that sometimes we input two images, but the model says it only sees one