app.py

# app.py — Gradio dengan decoding yang konsisten seperti di Colab
import os
import gradio as gr
import torch
from transformers import (
    AutoTokenizer, AutoModelForSeq2SeqLM, pipeline, GenerationConfig
)

MODEL_ID = os.getenv("MODEL_ID", "hasmar03/mt5_id2md")

# >>>>>> UBAH SESUAI DATA TRAINING KAMU <<<<<<
ID2MD_PREFIX = "translate Indonesian to Mandar: "
MD2ID_PREFIX = "translate Mandar to Indonesian: "
# Jika kamu melatih pakai token lain (mis. "id2md: " / "md2id: " atau ">>md<< "),
# ganti string di atas agar 100% sama.

# Default decoding (samakan dengan Colab)
DEFAULT_DECODE = dict(
    num_beams=5,
    length_penalty=1.0,
    no_repeat_ngram_size=3,
    early_stopping=True,
    max_new_tokens=128,
)

pipe = None
gen_cfg = None

def get_pipe():
    global pipe, gen_cfg
    if pipe is None:
        tok = AutoTokenizer.from_pretrained(MODEL_ID)
        mdl = AutoModelForSeq2SeqLM.from_pretrained(MODEL_ID)

        # Muat generation_config dari repo (jika ada)
        try:
            gen_cfg = GenerationConfig.from_pretrained(MODEL_ID)
            mdl.generation_config = gen_cfg
        except Exception:
            gen_cfg = mdl.generation_config  # fallback

        pipe = pipeline(
            "text2text-generation",
            model=mdl,
            tokenizer=tok,
            device=0 if torch.cuda.is_available() else -1,
        )
    return pipe

def build_prompt(text: str, direction: str):
    if direction == "Indonesia → Mandar":
        return f"{ID2MD_PREFIX}{text}"
    else:
        return f"{MD2ID_PREFIX}{text}"

def translate(text: str, direction: str,
              num_beams: int, max_new_tokens: int,
              no_repeat_ngram_size: int, length_penalty: float,
              do_sample: bool, temperature: float, top_p: float, top_k: int):
    p = get_pipe()
    prompt = build_prompt(text, direction)

    # Susun argumen generate; mulai dari DEFAULT_DECODE lalu override dari UI
    gen_args = dict(DEFAULT_DECODE)
    gen_args.update(
        num_beams=int(num_beams),
        max_new_tokens=int(max_new_tokens),
        no_repeat_ngram_size=int(no_repeat_ngram_size),
        length_penalty=float(length_penalty),
    )
    if do_sample:
        gen_args.update(do_sample=True, temperature=float(temperature),
                        top_p=float(top_p), top_k=int(top_k))
    else:
        gen_args.update(do_sample=False)

    out = p(prompt, **gen_args)[0]["generated_text"]
    return out

with gr.Blocks(title="Mandar ↔ Indonesia Translator") as demo:
    gr.Markdown("### Mandar ↔ Indonesia Translator")
    with gr.Row():
        direction = gr.Radio(
            ["Indonesia → Mandar", "Mandar → Indonesia"],
            value="Indonesia → Mandar", label="Arah"
        )
    src = gr.Textbox(label="Teks sumber", lines=3, placeholder="Ketik teks…")
    btn = gr.Button("Terjemahkan")
    out = gr.Textbox(label="Hasil", lines=3)

    with gr.Accordion("Advanced decoding", open=False):
        num_beams = gr.Slider(1, 10, value=DEFAULT_DECODE["num_beams"], step=1, label="num_beams")
        max_new_tokens = gr.Slider(16, 512, value=DEFAULT_DECODE["max_new_tokens"], step=8, label="max_new_tokens")
        no_repeat_ngram_size = gr.Slider(0, 10, value=DEFAULT_DECODE["no_repeat_ngram_size"], step=1, label="no_repeat_ngram_size")
        length_penalty = gr.Slider(0.0, 2.0, value=DEFAULT_DECODE["length_penalty"], step=0.1, label="length_penalty")
        do_sample = gr.Checkbox(False, label="Sampling (non-deterministic)")
        temperature = gr.Slider(0.1, 2.0, value=1.0, step=0.1, label="temperature")
        top_p = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="top_p")
        top_k = gr.Slider(0, 100, value=50, step=5, label="top_k")

    btn.click(
        translate,
        inputs=[src, direction, num_beams, max_new_tokens, no_repeat_ngram_size,
                length_penalty, do_sample, temperature, top_p, top_k],
        outputs=out,
        api_name="translate"
    )

demo.queue()

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