import gradio as gr
from huggingface_hub import InferenceClient

import torch

from transformers import AutoTokenizer
from model.modeling_llamask import LlamaskForCausalLM
from model.tokenizer_utils import generate_custom_mask, prepare_tokenizer

model_id = "meta-llama/Meta-Llama-3.1-8B-Instruct"
device = 'cpu'

model = LlamaskForCausalLM.from_pretrained(model_id, torch_dtype= torch.bfloat16)
model = model.to(device)
tokenizer = AutoTokenizer.from_pretrained(model_id, padding_side="left")

prepare_tokenizer(tokenizer)


def respond(
    message,
    history: list[tuple[str, str]],
    max_tokens,
    temperature,
):
    prompt = """<|start_header_id|>system<|end_header_id|>

    You are a helpful assistant.<|eot_id|><|start_header_id|>user<|end_header_id|>
    {message}
    <|eot_id|><|start_header_id|>assistant<|end_header_id|>
    """
    model_inputs = generate_custom_mask(tokenizer, [prompt], device)
   
    outputs = model.generate(temperature=0.7, max_tokens=64, **model_inputs)
    outputs = outputs[:, model_inputs['input_ids'].shape[1]:]
    result = tokenizer.batch_decode(outputs, skip_special_tokens=True)

    return result, []

"""
For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
"""
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Markdown("Please enter your message. Add privacy tags (<sensitive>...</sensitive>) around the words you want to hide. Only the most recent message submitted will be taken into account (no history is retained)."),
        gr.Slider(minimum=1, maximum=128, value=32, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
    ],
)


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