app.py

import gradio as gr
from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM

model_id = "MediaTek-Research/Breeze-7B-Instruct-v1_0"  # Breeze7 模型名稱（請確認 Huggingface 上正確名稱）

# 載入 tokenizer 和模型
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

generator = pipeline("text-generation", model=model, tokenizer=tokenizer, trust_remote_code=True)

def chat_fn(user_input, chat_history):
    chat_history = chat_history or []
    chat_history.append(f"你: {user_input}")

    # 取最近幾句聊天作為 prompt，避免太長
    prompt = "\n".join(chat_history) + "\nAI:"

    response = generator(prompt, max_new_tokens=100, temperature=0.7)
    reply = response[0]['generated_text'][len(prompt):].strip()

    chat_history.append(f"AI: {reply}")
    return reply, chat_history

with gr.Blocks() as demo:
    chat_history = gr.State([])

    chatbot = gr.Chatbot()
    user_input = gr.Textbox(show_label=False, placeholder="說點什麼...")

    user_input.submit(chat_fn, inputs=[user_input, chat_history], outputs=[chatbot, chat_history])

demo.launch()