app.py

import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig  
import torch
import os

# 預先定義 Hugging Face 模型
MODEL_NAMES = {
    "DeepSeek-R1-Distill-Qwen-7B": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
    "DeepSeek-R1-Distill-Llama-8B": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
}

HF_TOKEN = os.getenv("HF_TOKEN")

def load_model(model_path):
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True, token=HF_TOKEN)
    
    # 先載入 config，手動刪除量化設定，防止 FP8 問題
    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True, token=HF_TOKEN)
    if hasattr(config, "quantization_config"):
        del config.quantization_config  # 刪除量化配置，避免使用 FP8

    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        config=config,  # 使用已移除量化的 config
        trust_remote_code=True,
        token=HF_TOKEN,
        torch_dtype=torch.float16,  # 強制 FP16，避免 FP8
        device_map="auto",
    )
    return model, tokenizer


# 預設載入 DeepSeek-R1
current_model, current_tokenizer = load_model("deepseek-ai/DeepSeek-R1-Distill-Llama-8B")


def chat(message, history, model_name):
    """處理聊天訊息"""
    global current_model, current_tokenizer

    # 若模型不同則切換
    if model_name != current_model:
        current_model, current_tokenizer = load_model(model_name)

    device = "cuda" if torch.cuda.is_available() else "cpu"
    inputs = current_tokenizer(message, return_tensors="pt").to(device)
    outputs = current_model.generate(**inputs, max_length=1024)
    response = current_tokenizer.decode(outputs[0], skip_special_tokens=True)

    return response


with gr.Blocks() as app:
    gr.Markdown("## Chatbot with DeepSeek Models")

    with gr.Row():
        chat_interface = gr.ChatInterface(chat, streaming=True, save_history=True)
        model_selector = gr.Dropdown(
            choices=list(MODEL_NAMES.keys()), value="DeepSeek-R1-Distill-Llama-8B", label="Select Model"
        )

    chat_interface.append(model_selector)
    app.launch()