import gradio as gr import torch from transformers import AutoModelForCausalLM, AutoTokenizer # 使用 Qwen3.5-9B 量化版本(显存占用约 6GB) MODEL_ID = "Qwen/Qwen3.5-9B" print("正在加载模型...") tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( MODEL_ID, device_map="auto", trust_remote_code=True, dtype=torch.float16, ) print("模型加载完成!") def generate(text, max_length=512): if not text.strip(): return "请输入内容" messages = [{"role": "user", "content": text}] text_input = tokenizer.apply_chat_template( messages, tokenize=False, add_generation_prompt=True ) inputs = tokenizer(text_input, return_tensors="pt").to(model.device) outputs = model.generate( **inputs, max_new_tokens=max_length, do_sample=True, temperature=0.7, pad_token_id=tokenizer.eos_token_id, ) response = tokenizer.decode( outputs[0][len(inputs.input_ids[0]):], skip_special_tokens=True ) return response with gr.Blocks(title="Qwen 聊天", theme=gr.themes.Soft()) as demo: gr.Markdown("# 🤖 Qwen3.5-9B 在线聊天") input_text = gr.Textbox(label="输入", lines=3, placeholder="你好...") output_text = gr.Textbox(label="回复", lines=6) btn = gr.Button("发送", variant="primary") btn.click(fn=generate, inputs=input_text, outputs=output_text) if __name__ == "__main__": demo.launch()