huggingface spaces gradio run phi-4 model by transformers and @spaces.GPU

import spaces
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
import gradio as gr
import os
from threading import Thread

os.system("rm -rf /data-nvme/zerogpu-offload/*")

os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'

model = AutoModelForCausalLM.from_pretrained(
    "NyxKrage/Microsoft_Phi-4", 
    device_map="cuda", 
    torch_dtype="auto", 
    trust_remote_code=True, 
)
tokenizer = AutoTokenizer.from_pretrained("NyxKrage/Microsoft_Phi-4")

streamer = TextIteratorStreamer(tokenizer)

@spaces.GPU
def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
    seed,
):
    messages = [{"role": "system", "content": system_message}]

    for val in history:
        if val[0]:
            messages.append({"role": "user", "content": val[0]})
        if val[1]:
            messages.append({"role": "assistant", "content": val[1]})

    messages.append({"role": "user", "content": message})

    # Convert messages to the format expected by the model
    input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")

    torch.random.manual_seed(seed)

    generation_kwargs = dict(
        input_ids=input_ids,
        max_new_tokens=max_tokens,
        temperature=temperature,
        streamer=streamer,
        top_p=top_p,
        return_full_text=False,
        do_sample=True,
    )
    
    response = ""
    # Run the generation in a separate thread, so that we can fetch the generated text in a non-blocking way.
    thread = Thread(target=model.generate, kwargs=generation_kwargs)
    thread.start()
    # Print the generated text in real-time
    for new_text in streamer:
        response += new_text
    yield response
    
with gr.Blocks() as demo:
    with gr.Row():
        gr.LoginButton(min_width=100)
        gr.Markdown("""
                    This is the space I built.
                    As of 2025/1/7, this is the first phi-4 space.
                    If this helps you, and if you have enough money, can you give me 1$? I am facing a financial crisis.
                    If you do this, I will pass on the kindness.
                    This is my bank card number:5592921230414708
                    Thank you!!
                    """)
    
    with gr.Row():
        with gr.Column():
            system_message = gr.Textbox(value="You are a friendly Chatbot.", label="System message")
            max_tokens = gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens")
            temperature = gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature")
            top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)")
            seed = gr.Slider(minimum=0, maximum=20091114, value=42, step=1, label="Seed")
        
        with gr.Column():
            gr.ChatInterface(
                respond,
                additional_inputs=[system_message, max_tokens, temperature, top_p, seed],
            )


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