Update app.py

app.py

@@ -2,12 +2,10 @@ import spaces
 from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
 import gradio as gr
-import os
 
-os.environ['CUDA_VISIBLE_DEVICES'] = "0,1"  
+# ZeroGPU 环境会自动管理 GPU 分配，因此我们不设置 CUDA_VISIBLE_DEVICES
 USE_CUDA = torch.cuda.is_available()
-device_ids_parallel = [0]
-device = torch.device("cuda:{}".format(device_ids_parallel[0]) if USE_CUDA else "cpu")
+device = torch.device("cuda:0" if USE_CUDA else "cpu")
 
 # 初始化
 peft_model_id = "CMLM/ZhongJing-2-1_8b"
@@ -16,12 +14,25 @@ model = AutoModelForCausalLM.from_pretrained(base_model_id, device_map="auto")
 model.load_adapter(peft_model_id)
 tokenizer = AutoTokenizer.from_pretrained(
     "CMLM/ZhongJing-2-1_8b",
-    padding_side="right", 
+    padding_side="right",
     trust_remote_code=True,
     pad_token=''
 )
 
-#多轮  
+@spaces.GPU
+def single_turn_chat(question):
+    prompt = f"Question: {question}"
+    messages = [
+        {"role": "system", "content": "You are a helpful TCM medical assistant named 仲景中医大语言模型, created by 医哲未来 of Fudan University."},
+        {"role": "user", "content": prompt}
+    ]
+    input = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    model_inputs = tokenizer([input], return_tensors="pt").to(device)
+    generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
+    generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
+    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+    return response
+
 @spaces.GPU
 def multi_turn_chat(question, chat_history=None):
     if not isinstance(question, str):
@@ -32,51 +43,23 @@ def multi_turn_chat(question, chat_history=None):
     
     chat_history.append({"role": "user", "content": question})
     
-    # Apply the chat template and prepare the input 
     inputs = tokenizer.apply_chat_template(chat_history, tokenize=False, add_generation_prompt=True)
     model_inputs = tokenizer([inputs], return_tensors="pt").to(device)
-
-    try:
-        # Generate the response from the model
-        outputs = model.generate(model_inputs.input_ids, max_new_tokens=512)
-        generated_ids = outputs[:, model_inputs.input_ids.shape[-1]:]
-        response = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
-    except Exception as e:
-        raise RuntimeError("Error in model generation: " + str(e))
-    
-    # Append the assistant's response to the chat history
-    chat_history.append({"role": "assistant", "content": response})
     
-    # Format the chat history for output
-    tempass = ""
-    tempuser = ""
-    formatted_history = []
-    for entry in chat_history:
-        if entry['role'] == 'user':
-            tempuser = entry['content'] 
-        elif entry['role'] == 'assistant':
-            tempass = entry['content']
-            temp = tempuser,tempass 
-            formatted_history.append(temp)
+    outputs = model.generate(model_inputs.input_ids, max_new_tokens=512)
+    generated_ids = outputs[:, model_inputs.input_ids.shape[-1]:]
+    response = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
     
-    return formatted_history, chat_history
-
-
-def clear_history():
-    return [], []
+    chat_history.append({"role": "assistant", "content": response})
+    return chat_history
 
-# 多轮界面  
-with gr.Blocks() as multi_turn_interface:
-    chatbot = gr.Chatbot(label="仲景GPT-V2-1.8B 多轮对话")
-    state = gr.State([])  
-    with gr.Row():
-        with gr.Column(scale=6):
-            user_input = gr.Textbox(label="输入",placeholder="输入你的问题")
-        with gr.Column(scale=1):
-            submit_btn = gr.Button("提交")
-            clear_history_btn = gr.Button("清除历史对话")
-    submit_btn.click(multi_turn_chat, [user_input, state], [chatbot, state], concurrency_limit=10)
-    clear_history_btn.click(fn=clear_history, inputs=None, outputs=[chatbot, state], queue=False)
-    user_input.submit(multi_turn_chat, [user_input, state], [chatbot, state], concurrency_limit=10)
+# 单轮界面
+single_turn_interface = gr.Interface(
+    fn=single_turn_chat,
+    inputs=["text"],
+    outputs="text",
+    title="仲景GPT-V2-1.8B 单轮对话",
+    description="Unlocking the Wisdom of Traditional Chinese Medicine with AI."
+)
 
-multi_turn_interface.launch(debug=True, server_name='0.0.0.0', server_port=6006, max_threads=200)
+# 多轮界面配置与之前保持一致