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)

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
)

def parse_llama_guard_output(result):
    lines = [line.strip().lower() for line in result.split('\n') if line.strip()]
    
    if not lines:
        return "Error", "No valid output", result

    safety_status = next((line for line in lines if line in ['safe', 'unsafe']), None)
    
    if safety_status == 'safe':
        return "Safe", "None", result
    elif safety_status == 'unsafe':
        violated_categories = next((lines[i+1] for i, line in enumerate(lines) if line == 'unsafe' and i+1 < len(lines)), "Unspecified")
        return "Unsafe", violated_categories, result
    else:
        return "Error", f"Invalid output: {safety_status}", result

@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)
    
    with torch.no_grad():
        output = model.generate(
            input_ids=input_ids, 
            max_new_tokens=100,
            pad_token_id=tokenizer.eos_token_id,
        )
    
    prompt_len = input_ids.shape[-1]
    result = tokenizer.decode(output[0][prompt_len:], skip_special_tokens=True)
    
    return parse_llama_guard_output(result)

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

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