qwen2.5-vl model infer has an compute hash_feature function cost 84ms，and the qwen-vl vit forward cost 53ms, so it is necessary? or has another function to replace it?

[TZHelloWorld]

env：

• sglang: v0.4.6.post5
• GPU: NVIDIA GeForce RTX 4090

Start sglang service：

export model_name="Qwen/Qwen2.5-VL-3B-Instruct"
export SGLANG_TORCH_PROFILER_DIR=/root/sglang # sglang proiler dir
gpu_devices="3"
CUDA_VISIBLE_DEVICES=${gpu_devices} nohup python3 -m sglang.launch_server \
    --model-path ${model_name} \
    --tokenizer-path ${model_name} \
    --served-model-name base_model --host 0.0.0.0 --port 20001 --trust-remote-code \
    --mem-fraction-static 0.6 \
    --dp-size 1 \
    --chat-template "qwen2-vl" \
    --disable-custom-all-reduce \
    --chunked-prefill-size  -1 \
    --cuda-graph-max-bs 12 \
    --enable-torch-compile >/tmp/log_one.txt 2>&1&

and then send request:

# start trace
curl -X POST http://10.38.xxx.xxx:20001/start_profile -H "Content-Type: application/json"

# sed request （twice or more）
curl -X POST http://xxx.xxx.xxx.xxx:xx/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
  "model": "model",
  "finish_reason": "inf_loop",
  "stream":true,
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "在下面图片中是什么？"
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"
          }
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"
          }
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"
          }
        },
        {
          "type": "image_url",
          "image_url": {
            "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/svd/rocket.png"
          }
        }
      ]
    }
  ]
}'

# stop trace
curl -X POST http://10.38.xxx.xxx:20001/stop_profile -H "Content-Type: application/json"

it can trace a *.json.gz, with open in https://ui.perfetto.dev/，and then find python/sglang/srt/managers/schedule_batch.py compute hash_feature function cost 84ms, it is necessnary？or has any other function to replace it？

because the qwen-vl model vit forwad cost 53ms:

[JustinTong0323]

#8179