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 = "google/gemma-2-9b-it"

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


@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)

    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."



def call_generate_text(message, history):
    history.append({"role": "assistant", "content": message})
    print(message)
    print(history)
   
    #messages = history + message
   # messages = [{"role":"user","content":message}]
    try:
        text = generate_text(history)
        history.append({"role": "assistant", "content": text})
        return text
    except RuntimeError  as e:
        print(f"An unexpected error occurred: {e}")
       
    return ""

demo = gr.ChatInterface(call_generate_text)

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