app.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
from threading import Thread
import os

model_id = "novapixelentretaiment/Lumin-Haiku-4-Instruct"
# Archivo corregido para coincidir con el nombre oficial del repo en HF
gguf_file = "qwen2.5-coder-3b-instruct-q4_k_m.gguf" 
token = os.environ.get("HF_TOKEN")

print("Cargando Lumin Nano 2.1 (GGUF Optimized)...")


tokenizer = AutoTokenizer.from_pretrained(model_id, token=token, trust_remote_code=True)


model = AutoModelForCausalLM.from_pretrained(
    model_id,
    gguf_file=gguf_file,
    token=token,
    torch_dtype=torch.float32, # CPU friendly
    trust_remote_code=True
)

def parse_thought(text):
    if "<think>" in text:
        if "</think>" in text:
            parts = text.split("</think>")
            return f"Pensamiento: {parts[0].replace('<think>', '').strip()}\n\nRespuesta: {parts[1].strip()}"
        else:
            return f"Pensamiento: {text.replace('<think>', '').strip()}"
    return text

def chat_stream(message, history, system_message, max_tokens, temperature, top_p):
    messages = [{"role": "system", "content": system_message}]
    for h in history:
        if h[0]: messages.append({"role": "user", "content": h[0]})
        if h[1]: messages.append({"role": "assistant", "content": h[1]})
    messages.append({"role": "user", "content": message})

    text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer([text], return_tensors="pt").to(model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
    
    generate_kwargs = dict(
        **inputs,
        streamer=streamer,
        max_new_tokens=max_tokens,
        temperature=temperature,
        top_p=top_p,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=[tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|im_end|>")],
    )

    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    partial_text = ""
    for new_text in streamer:
        partial_text += new_text
        if "<|im_end|>" in partial_text:
            partial_text = partial_text.split("<|im_end|>")[0]
            yield parse_thought(partial_text)
            break
        yield parse_thought(partial_text)

with gr.Blocks(title="Lumin Nano 2.1") as demo:
    gr.Markdown("Lumin Nano 2.1 - Spanish Only")
    
    gr.ChatInterface(
        chat_stream,
        additional_inputs=[
            gr.Textbox(value="Eres Lumin Nano 2.1. UNICAMENTE puedes pensar y responder en ESPAÑOL. Tienes PROHIBIDO usar el inglés. Sé directo, conciso y nunca uses emojis.", label="System Message"),
            gr.Slider(1, 1024, 256, label="Max Tokens"),
            gr.Slider(0.01, 1.0, 0.1, label="Temperature"),
            gr.Slider(0.1, 1.0, 0.9, label="Top-p"),
        ],
    )

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