app.py

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

# === Carregar modelo local (CPU) ===
model_name = "lambdaindie/lambda-1v-1B"  # troque pelo teu
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to("cpu")  # <- CPU aqui

# === Streamer global para interrupção ===
stop_signal = {"stop": False}

def generate_stream(prompt, max_tokens=512, temperature=0.7, top_p=0.95):
    stop_signal["stop"] = False
    inputs = tokenizer(prompt, return_tensors="pt").to("cpu")

    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)

    generation_thread = threading.Thread(
        target=model.generate,
        kwargs=dict(
            input_ids=inputs["input_ids"],
            attention_mask=inputs["attention_mask"],
            streamer=streamer,
            max_new_tokens=max_tokens,
            do_sample=True,
            temperature=temperature,
            top_p=top_p,
            pad_token_id=tokenizer.eos_token_id,
        )
    )
    generation_thread.start()

    output = ""
    for token in streamer:
        if stop_signal["stop"]:
            break
        output += token
        yield output.strip()

def stop_stream():
    stop_signal["stop"] = True

def respond(message, history, system_message, max_tokens, temperature, top_p):
    messages = [{"role": "system", "content": system_message}] if system_message else []

    for user, assistant in history[-3:]:  # Limita a 3 interações passadas
        if user:
            messages.append({"role": "user", "content": user})
        if assistant:
            messages.append({"role": "assistant", "content": assistant})

    thinking_prompt = messages + [{"role": "user", "content": f"{message}\n\nThink step-by-step before answering."}]
    thinking_text = "\n".join([f"{m['role']}: {m['content']}" for m in thinking_prompt])

    reasoning = ""
    yield '<div class="markdown-think">Thinking...</div>'

    start = time.time()
    for token in generate_stream(thinking_text, max_tokens, temperature, top_p):
        reasoning = token
        yield f'<div class="markdown-think">{reasoning.strip()}</div>'

    elapsed = time.time() - start
    yield f"""
    <div style="margin-top:12px;padding:8px 12px;background-color:#222;border-left:4px solid #888;
            font-family:'JetBrains Mono', monospace;color:#ccc;font-size:14px;">
        Pensou por {elapsed:.1f} segundos
    </div>
    """

    final_prompt = thinking_text + f"\n\nuser: {message}\nassistant: {reasoning.strip()}\nuser: Now answer based on your reasoning above.\nassistant:"
    final_answer = ""

    for token in generate_stream(final_prompt, max_tokens, temperature, top_p):
        final_answer = token
        yield final_answer.strip()

# === Interface ===

css = """
@import url('https://fonts.googleapis.com/css2?family=JetBrains+Mono&display=swap');
* { font-family: 'JetBrains Mono', monospace !important; }
html, body, .gradio-container {
    background-color: #111 !important;
    color: #e0e0e0 !important;
}
textarea, input, button, select {
    background-color: transparent !important;
    color: #e0e0e0 !important;
    border: 1px solid #444 !important;
}
.markdown-think {
    background-color: #1e1e1e;
    border-left: 4px solid #555;
    padding: 10px;
    margin-bottom: 8px;
    font-style: italic;
    white-space: pre-wrap;
    animation: pulse 1.5s infinite ease-in-out;
}
@keyframes pulse {
    0% { opacity: 0.6; }
    50% { opacity: 1.0; }
    100% { opacity: 0.6; }
}
"""

theme = gr.themes.Base(
    primary_hue="gray",
    font=[gr.themes.GoogleFont("JetBrains Mono"), "monospace"]
).set(
    body_background_fill="#111",
    body_text_color="#e0e0e0",
    input_background_fill="#222",
    input_border_color="#444",
    button_primary_background_fill="#333",
    button_primary_text_color="#e0e0e0",
)

chatbot = gr.ChatInterface(
    fn=respond,
    title="λambdAI",
    css=css,
    theme=theme,
    additional_inputs=[
        gr.Textbox(value="", label="System Message"),
        gr.Slider(64, 2048, value=512, step=1, label="Max Tokens"),
        gr.Slider(0.1, 2.0, value=0.7, step=0.1, label="Temperature"),
        gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p")
    ]
)

stop_btn = gr.Button("Parar Geração")
stop_btn.click(fn=stop_stream, inputs=[], outputs=[])

app = gr.Blocks()
with app:
    chatbot.render()
    stop_btn.render()

if __name__ == "__main__":
    app.launch(share=True)