app.py

import spaces
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import os

if not os.path.exists("./phi-4"):
    os.system('huggingface-cli download matteogeniaccio/phi-4 --local-dir ./phi-4 --include "phi-4/*"')

from safetensors.torch import load_file, save_file

@spaces.GPU
def merge_safetensors(input_dir, output_file):
    # 获取所有分片文件
    files = sorted([f for f in os.listdir(input_dir) if f.startswith('model-') and f.endswith('.safetensors')])
    
    # 合并所有张量
    merged_state_dict = {}
    for file in files:
        file_path = os.path.join(input_dir, file)
        print(f"Loading {file}...")
        state_dict = load_file(file_path)
        merged_state_dict.update(state_dict)
    
    # 保存合并后的文件
    print(f"Saving merged model to {output_file}...")
    save_file(merged_state_dict, output_file)
    print("Done!")

# 使用示例
input_dir = "./phi-4/phi-4"  # 包含分片文件的目录
output_file = "./phi-4/phi-4/model.safetensors"  # 合并后的文件路径

if not os.path.exists(output_file):
    merge_safetensors(input_dir, output_file)

# 加载 phi-4 模型和 tokenizer
torch.random.manual_seed(0)

model = AutoModelForCausalLM.from_pretrained(
    "./phi-4/phi-4",  # 模型路径
    device_map="cuda",  # 使用 GPU
    # torch_dtype="auto",  # 自动选择数据类型
    trust_remote_code=True,  # 允许远程代码加载
    use_safetensors=True,
    torch_dtype=torch.float16,
)
tokenizer = AutoTokenizer.from_pretrained("./phi-4/phi-4")

# 设置 pipeline

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

# 响应函数
@spaces.GPU
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    # 构造消息内容
    messages = [{"role": "system", "content": system_message}]
    for user_msg, assistant_msg in history:
        if user_msg:
            messages.append({"role": "user", "content": user_msg})
        if assistant_msg:
            messages.append({"role": "assistant", "content": assistant_msg})
    messages.append({"role": "user", "content": message})

    # 将消息转换为字符串格式（适用于 text-generation）
    input_text = "\n".join(
        f"{msg['role']}: {msg['content']}" for msg in messages
    )

    # 生成响应
    generation_args = {
        "max_new_tokens": max_tokens,
        "temperature": temperature,
        "top_p": top_p,
        "do_sample": temperature > 0,
        "return_full_text": False,
    }
    output = pipe(input_text, **generation_args)
    response = output[0]["generated_text"]

    # 返回流式响应
    for token in response:
        yield token

# Gradio 界面
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)"
        ),
    ],
)

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