app.py

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

huggingface_token = os.getenv('HUGGINGFACE_TOKEN')
if not huggingface_token:
    raise ValueError("HUGGINGFACE_TOKEN environment variable is not set")

model_id = "meta-llama/Llama-Guard-3-8B-INT8"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.bfloat16

quantization_config = BitsAndBytesConfig(load_in_8bit=True)

@spaces.GPU
def load_model():
    tokenizer = AutoTokenizer.from_pretrained(model_id, token=huggingface_token)
    model = AutoModelForCausalLM.from_pretrained(
        model_id, 
        torch_dtype=dtype, 
        device_map="auto",
        quantization_config=quantization_config,
        token=huggingface_token,
        low_cpu_mem_usage=True
    )
    return tokenizer, model

tokenizer, model = load_model()

@spaces.GPU
def moderate(user_input, assistant_response):
    chat = [
        {"role": "user", "content": user_input},
        {"role": "assistant", "content": assistant_response},
    ]
    input_ids = tokenizer.apply_chat_template(chat, return_tensors="pt").to(device)
    output = model.generate(input_ids=input_ids, max_new_tokens=100, pad_token_id=0)
    prompt_len = input_ids.shape[-1]
    return tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)

iface = gr.Interface(
    fn=moderate,
    inputs=[
        gr.Textbox(lines=3, label="User Input"),
        gr.Textbox(lines=3, label="Assistant Response")
    ],
    outputs=gr.Textbox(label="Moderation Result"),
    title="Llama Guard Moderation",
    description="Enter a user input and an assistant response to check for content moderation."
)

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