import gradio as gr
import time
from optimum.intel import OVModelForCausalLM
from transformers import AutoTokenizer, pipeline

# Load the model and tokenizer
model_id = "hsuwill000/DeepSeek-R1-Distill-Qwen-1.5B-openvino"
model = OVModelForCausalLM.from_pretrained(model_id, device="CPU")  # 明确指定设备
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Create generation pipeline
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)

def respond(message):
    try:
        start_time = time.time()
        
        # 強化 Prompt 讓模型輸出更合理
        instruction = (
            "請用簡單、準確的語言回答問題，避免冗長和重複內容。\n"
            "User: " + message + "\n"
            "Assistant: "
        )

        # Generate response with improved settings
        response = pipe(
            instruction,
            max_length=200,  # 限制最大輸出長度，防止無限重複
            truncation=True,
            num_return_sequences=1,
            temperature=0.3,  # 保持一定創意但減少胡言亂語
            top_p=0.8,  # 增加答案多樣性，減少無意義重複
            repetition_penalty=1.5,  # 降低重複字詞的機率
        )
        generated_text = response[0]['generated_text'].strip()

        # 提取 "Assistant:" 之後的部分
        if "Assistant:" in generated_text:
            reply = generated_text.split("Assistant:")[-1].strip()
        else:
            reply = generated_text
        
        inference_time = time.time() - start_time
        print(f"Inference time: {inference_time:.4f} seconds")

        return [(message, reply)]
    
    except Exception as e:
        print(f"Error: {e}")
        return [(message, "Sorry, something went wrong. Please try again.")]

# Set up Gradio chat interface
with gr.Blocks() as demo:
    gr.Markdown("# DeepSeek-R1-Distill-Qwen-1.5B-openvino Chat")
    gr.Markdown("Chat with DeepSeek-R1-Distill-Qwen-1.5B-openvino model.")
    
    chatbot = gr.Chatbot()
    msg = gr.Textbox(label="Your Message")
    
    msg.submit(respond, msg, chatbot)

if __name__ == "__main__":
    demo.launch(share=True)