app.py

import os
from dotenv import load_dotenv
import streamlit as st
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Carregar as variáveis de ambiente do arquivo .env
load_dotenv()

# Recuperar o token da variável de ambiente
hf_token = os.getenv("HF_API_TOKEN")

# Carregando o modelo e o tokenizer
@st.cache_resource
def load_model():
    model_name = "meta-llama/Meta-Llama-3.1-8B-Instruct"
    tokenizer = AutoTokenizer.from_pretrained(model_name, use_auth_token=hf_token)
    model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", torch_dtype=torch.float16, use_auth_token=hf_token)
    return model, tokenizer

model, tokenizer = load_model()

st.title("Chatbot com Meta-Llama 3.1")

# Entrada do usuário
user_input = st.text_input("Você:", "")

if user_input:
    # Tokenizando a entrada do usuário
    inputs = tokenizer(user_input, return_tensors="pt").to("cuda")

    # Gerando a resposta
    with torch.no_grad():
        outputs = model.generate(inputs.input_ids, max_length=150, do_sample=True, top_p=0.95, temperature=0.7)

    # Decodificando e mostrando a resposta
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    st.text_area("Resposta do Chatbot:", value=response, height=200)