app.py

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

huggingface_token = os.getenv("HUGGINGFACE_TOKEN")
if not huggingface_token:
    pass
    print("no HUGGINGFACE_TOKEN if you need set secret ")
    #raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")

model_id = "microsoft/Phi-3-mini-128k-instruct"

device = "auto" # torch.device("cuda" if torch.cuda.is_available() else "cpu")
dtype = torch.bfloat16

tokenizer = AutoTokenizer.from_pretrained(model_id)#, token=huggingface_token)


import time
time.sleep(10)


print(model_id,device,dtype)
histories = []
contents = []
    
def call_generate_text(prompt, system_message="You are a helpful assistant."):
    
    print(histories)
    print(contents)
    
    if prompt =="":
        print("empty prompt return")
        return ""
    global initialized
    if not initialized:
        initialized = True
        #return
    try:
        text = generate_text(prompt,system_message)
        contents.append(text)
        return text
    except RuntimeError  as e:
        print(f"An unexpected error occurred: {e}")

    return ""

    
        
initialized = False

iface = gr.Interface(
    fn=call_generate_text,
    inputs=[
        gr.Textbox(lines=3, label="Input Prompt"),
        gr.Textbox(lines=2, label="System Message", value="あなたは親切なアシスタントで常に日本語で返答します。"),
    ],
    outputs=gr.Textbox(label="Generated Text"),
    title="Phi-3-mini-128k-instruct",
    description="Phi-3-mini-128k-instruct",
)
print("Initialized")

# keeping model seems make crash

@spaces.GPU(duration=100)
def generate_text(prompt, system_message="You are a helpful assistant."):
    #print(prompt,system_message)

    global histories
   
    model = AutoModelForCausalLM.from_pretrained(
        model_id ,torch_dtype=dtype,device_map=device # token=huggingface_token
    )
    #print(system_message)
    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer,torch_dtype=dtype,device_map=device) #pipeline has not to(device)

    messages = [
        {"role": "system", "content": system_message},
    ]
    
    messages += histories 

    user_message = {"role": "user", "content": prompt}

    messages += [user_message]

    #print(messages)
    
    result = text_generator(messages, max_new_tokens=256, do_sample=True, temperature=0.7)

    generated_output = result[0]["generated_text"]
    if isinstance(generated_output, list):
        for message in reversed(generated_output):
            if message.get("role") == "assistant":
                content= message.get("content", "No content found.")
                histories += [user_message,{"role": "assistant", "content": content}]
                print(f"history = {len(histories)}")
                return content
            
        return "No assistant response found."
    else:
        return "Unexpected output format."

if __name__ == "__main__":
    print("Main")
    iface.launch()