import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Carrega tokenizer e modelo manualmente
tokenizer = AutoTokenizer.from_pretrained("lambdaindie/lambda-1v-1B")
model = AutoModelForCausalLM.from_pretrained("lambdaindie/lambda-1v-1B")

# Garante uso de CPU
device = torch.device("cpu")
model.to(device)

# Função de geração
def responder(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=50,
        do_sample=True,
        top_p=0.95,
        temperature=0.8,
        pad_token_id=tokenizer.eos_token_id  # evita warning se for causal
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

# Interface Gradio
iface = gr.Interface(
    fn=responder,
    inputs=gr.Textbox(lines=2, placeholder="Escreve algo..."),
    outputs="text",
    title="Lambda-1v-1B",
    description="Modelo local de geração de texto criado por Marius Jabami.",
)

iface.launch()