app.py

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

# =========================
# CONFIG
# =========================
MODEL_ID = "AxionLab-Co/DogeAI-v2.0-4B-Reasoning"
MAX_NEW_TOKENS = 256  # menor = menos timeout em CPU

tokenizer = None
model = None


# =========================
# LOAD MODEL (LAZY + SAFE)
# =========================
def load_model():
    global tokenizer, model

    if model is None:
        tokenizer = AutoTokenizer.from_pretrained(
            MODEL_ID,
            use_fast=True
        )

        model = AutoModelForCausalLM.from_pretrained(
            MODEL_ID,
            device_map="cpu",
            torch_dtype=torch.float32,
            low_cpu_mem_usage=True
        )

        model.eval()

    return tokenizer, model


# =========================
# PROMPT (CPU-FRIENDLY)
# =========================
def build_prompt(user_input: str) -> str:
    return f"""You are DogeAI-v2.0-4B-Reasoning.

Think step by step internally.
Do not reveal your full chain-of-thought.
Provide a clear final answer with a short explanation.

If the user speaks Brazilian Portuguese:
- use Brazilian slang lightly
- keep the Doge vibe 🐕🇧🇷
- stay serious and logical

User:
{user_input}

Assistant:
"""


# =========================
# CHAT FUNCTION (SSE-SAFE)
# =========================
def chat(user_input):
    tokenizer, model = load_model()

    # mantém o SSE vivo imediatamente
    yield "🤔 DogeAI está pensando... segura aí..."

    prompt = build_prompt(user_input)

    inputs = tokenizer(
        prompt,
        return_tensors="pt"
    )

    with torch.no_grad():
        output = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            temperature=0.7,
            top_p=0.9,
            do_sample=True
        )

    text = tokenizer.decode(
        output[0],
        skip_special_tokens=True
    )

    # remove o prompt da resposta final
    response = text.split("Assistant:", 1)[-1].strip()

    yield response


# =========================
# GRADIO UI
# =========================
with gr.Blocks(title="DogeAI-v2.0-4B-Reasoning") as demo:
    gr.Markdown(
        "# 🐕 DogeAI-v2.0-4B-Reasoning\n"
        "**4B reasoning model rodando em CPU no HF Space**\n\n"
        "Pensamento explícito interno, resposta clara externa."
    )

    input_box = gr.Textbox(
        label="Pergunta",
        placeholder="Pergunta que exige raciocínio de verdade...",
        lines=4
    )

    output_box = gr.Textbox(
        label="Resposta do DogeAI",
        lines=14
    )

    run_btn = gr.Button("Pensar 🧠🐕")

    run_btn.click(
        fn=chat,
        inputs=input_box,
        outputs=output_box
    )

demo.launch()