app.py

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


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 = "Qwen/Qwen1.5-0.5B-Chat"
model_id = "Kendamarron/Tokara-0.5B-Chat-v0.1"
model_id = "Qwen/Qwen2-0.5B-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)

print(model_id,device,dtype)
histories = []
#model = None


def call_generate_text(prompt, system_message="You are a helpful assistant."):
    if prompt =="":
        print("empty prompt return")
        return ""

    global histories
    #global model
    #if model != None:# and model.is_cuda:
    #    print("Model is alive")
    #else:
    #    model = AutoModelForCausalLM.from_pretrained(
    #    model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
    #)
        
    messages = [
        {"role": "system", "content": system_message},
    ]
    
    messages += histories 

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

    messages += [user_message]
    
    try:
        text = generate_text(messages)
        histories += [user_message,{"role": "assistant", "content": text}]
        #model.to("cpu")
        return text
    except RuntimeError  as e:
        print(f"An unexpected error occurred: {e}")
        #model = None

    return ""

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=f"{model_id}",
    description=f"{model_id} CPU",
)
print("Initialized")

@spaces.GPU(duration=120)
def generate_text(messages):

    model = AutoModelForCausalLM.from_pretrained(
        model_id, token=huggingface_token ,torch_dtype=dtype,device_map=device
    )
        

    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer,torch_dtype=dtype,device_map=device) #pipeline has not to(device)
    result = text_generator(messages, max_new_tokens=256, do_sample=True, temperature=0.7,repetition_penalty=1.1,top_p=0.95,top_k=40)

    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.")
                return content
            
        return "No assistant response found."
    else:
        return "Unexpected output format."

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