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
    raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")

model_id = "google/gemma-2-9b-it"
#model_id = "microsoft/Phi-3-mini-128k-instruct"
# device_map style value auto not cuda
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)
@spaces.GPU
def generate_text(prompt, system_message="You are a helpful assistant."):
    model = AutoModelForCausalLM.from_pretrained(
        model_id, torch_dtype=dtype,device_map=device, token=huggingface_token
    )
    text_generator = pipeline("text-generation", model=model, tokenizer=tokenizer, torch_dtype=dtype, device_map=device)

    messages = [
        {"role": "system", "content": system_message},
        {"role": "user", "content": prompt},
    ]

    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":
                return message.get("content", "No content found.")
        return "No assistant response found."
    else:
        return "Unexpected output format."


iface = gr.Interface(
    fn=generate_text,
    inputs=[
        gr.Textbox(lines=3, label="Input Prompt"),
        gr.Textbox(lines=2, label="System Message", value="You are a helpful assistant."),
    ],
    outputs=gr.Textbox(label="Generated Text"),
    title="google/gemma-2-9b-it Text Generation",
    description="Enter a prompt and optional system message to generate text using the google/gemma-2-9b-it model.",
)

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