阿里千问大模型(Qwen2.5-VL-7B-Instruct)部署

news2025/3/12 20:25:44

参考链接

知乎帖子
B站视频
huggingface 镜像网站（不太全，比如 Qwen/Qwen2.5-VL-7B-Instruct就没有）
huggingface 5种下载方式汇总
通过huggingface-cli下载模型

在这里插入图片描述

不一样的部分是预训练权重的下载和demo

首先安装huggingface_hub

pip install -U huggingface_hub

设置镜像

export HF_ENDPOINT=https://hf-mirror.com

windows端需要添加系统变量。

名称：HF_ENDPOINT，值： "https://hf-mirror.com"

然后通过huggingface-cli下载模型，

huggingface-cli download --resume-download  Qwen/Qwen2.5-VL-7B-Instruct --local-dir ./ --local-dir-use-symlinks False --resume-download

参考：通过huggingface-cli下载模型

运行DEMO

加载模型方式

如果希望下载到指定的目录，可以往from_pretrained方法传入cache_dir 参数，如下所示：

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("internlm/internlm2-chat-7b", torch_dtype=torch.float16, trust_remote_code=True, cache_dir='/home/{username}/huggingface').cuda()

运行以下代码

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

# 加载模型和处理器
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 "请上传一张图片。"
    
    # 构建消息格式
    messages = [
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "image": image,  # Gradio将自动处理图片路径
                },
                {"type": "text", "text": text_prompt if text_prompt else "Describe this image."},
            ],
        }
    ]
    
    try:
        # 准备推理输入
        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)

        # 生成输出
        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"处理过程中出现错误: {str(e)}"

# 创建Gradio界面
with gr.Blocks() as demo:
    gr.Markdown("# Qwen2.5-VL 图像理解演示")
    
    with gr.Row():
        with gr.Column():
            image_input = gr.Image(type="filepath", label="上传图片")
            text_input = gr.Textbox(
                placeholder="请输入提示语（如不输入，默认描述图片）", 
                label="提示语"
            )
            submit_btn = gr.Button("提交")
        
        with gr.Column():
            output = gr.Textbox(label="输出结果")
    
    submit_btn.click(
        fn=process_image_and_text,
        inputs=[image_input, text_input],
        outputs=output
    )

    gr.Examples(
        examples=[
            ["path/to/example1.jpg", "这张图片里有什么？"],
            ["path/to/example2.jpg", "描述图中的场景"],
        ],
        inputs=[image_input, text_input],
    )

# 启动应用
if __name__ == "__main__":
    demo.launch(share=True)