import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel, PeftConfig
import torch

# Carrega config do adaptador
config = PeftConfig.from_pretrained("lambdaindie/lambda-1v-1B")

# Carrega modelo base + LoRA
base_model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path)
model = PeftModel.from_pretrained(base_model, "lambdaindie/lambda-1v-1B")

# Tokenizer
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Envia pra CPU
device = torch.device("cpu")
model.to(device)

def responder(prompt):
    inputs = tokenizer(prompt, return_tensors="pt").to(device)
    outputs = model.generate(
        **inputs,
        max_new_tokens=50,
        temperature=0.8,
        top_p=0.95,
        pad_token_id=tokenizer.eos_token_id,
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

iface = gr.Interface(fn=responder,
                     inputs=gr.Textbox(lines=2, placeholder="Escreve algo..."),
                     outputs="text",
                     title="Lambda-1v-1B (LoRA)",
                     description="Modelo LoRA fine-tuned por Marius Jabami.")
iface.launch()