(qwen_vl) Ubuntu@0017-dsm-prxmx30138:~$ history 20
    8  conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
    9  pip install git+https://github.com/huggingface/transformers accelerate
   10  pip install beautifulsoup4 tinycss2
   11  pip install six
   12  pip install "qwen-vl-utils[decord]==0.0.8"
   13  pip install git+https://github.com/huggingface/transformers accelerate
   14  clear
   15  pip install "qwen-vl-utils[decord]==0.0.8"
   16  clear
   17  nano app.py
   18  python app.py
   19  clear
   20  python app.py
   21  pip install gradio
   22  clear
   23  python app.py
   24  clear
   25  history 20  # 顯示最近20條命令

import gradio as gr
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
import torch

# 加載模型和處理器
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-7B-Instruct", 
    torch_dtype="auto", 
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")

def process_image_and_text(image, text_prompt):
    if image is None:
        return "請上傳一張圖片。"
    
    # 構(gòu)建消息格式
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": image,  # Gradio將自動處理圖片路徑
                },
                {"type": "text", "text": text_prompt if text_prompt else "Describe this image."},
            ],
        }
    ]
    
    try:
        # 準備推理輸入
        text = processor.apply_chat_template(
            messages, tokenize=False, add_generation_prompt=True
        )
        image_inputs, video_inputs = process_vision_info(messages)
        inputs = processor(
            text=[text],
            images=image_inputs,
            videos=video_inputs,
            padding=True,
            return_tensors="pt",
        )
        inputs = inputs.to(model.device)

        # 生成輸出
        with torch.no_grad():
            generated_ids = model.generate(**inputs, max_new_tokens=128)
            generated_ids_trimmed = [
                out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
            ]
            output_text = processor.batch_decode(
                generated_ids_trimmed, 
                skip_special_tokens=True, 
                clean_up_tokenization_spaces=False
            )
        
        return output_text[0]
    
    except Exception as e:
        return f"處理過程中出現(xiàn)錯誤: {str(e)}"

# 創(chuàng)建Gradio界面
with gr.Blocks() as demo:
    gr.Markdown("# Qwen2.5-VL 圖像理解演示")
    
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(type="filepath", label="上傳圖片")
            text_input = gr.Textbox(
                placeholder="請輸入提示語（如不輸入，默認描述圖片）", 
                label="提示語"
            )
            submit_btn = gr.Button("提交")
        
        with gr.Column():
            output = gr.Textbox(label="輸出結(jié)果")
    
    submit_btn.click(
        fn=process_image_and_text,
        inputs=[image_input, text_input],
        outputs=output
    )

    gr.Examples(
        examples=[
            ["path/to/example1.jpg", "這張圖片里有什么？"],
            ["path/to/example2.jpg", "描述圖中的場景"],
        ],
        inputs=[image_input, text_input],
    )

# 啟動應(yīng)用
if __name__ == "__main__":
    demo.launch(share=True)