app.py

import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, MarianMTModel, MarianTokenizer
import shutil, os
shutil.rmtree(os.path.expanduser("~/.cache/huggingface"), ignore_errors=True)
shutil.rmtree(os.path.expanduser("~/.cache/torch"), ignore_errors=True)

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")

MODEL_OPTIONS = [
    "Helsinki-NLP (Tira ondo)", # Round-trip OPUS-MT en→es→en
    "FLAN-T5-base (Google gaizki xamar)"
]

# Cache
CACHE = {}

# --- FLAN loader (Google-style Euskera correction) ---
def load_flan():
    if "flan" not in CACHE:
        tok = AutoTokenizer.from_pretrained("google/flan-t5-base")
        mdl = AutoModelForSeq2SeqLM.from_pretrained(
            "google/flan-t5-base",
            low_cpu_mem_usage=True,
            torch_dtype="auto"
        ).to(DEVICE)
        CACHE["flan"] = (mdl, tok)
    return CACHE["flan"]

def run_flan(sentence: str) -> str:
    model, tok = load_flan()
    prompt = f"Euskara zuzen gramatikalki eta idatzi modu naturalean: {sentence}"
    inputs = tok(prompt, return_tensors="pt").to(DEVICE)
    with torch.no_grad():
        out = model.generate(**inputs, max_new_tokens=96, num_beams=4)
    return tok.decode(out[0], skip_special_tokens=True).strip()

# --- Euskera round-trip loader ---
def load_euskera():
    if "eus" not in CACHE:
        tok1 = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-eu-es")
        mdl1 = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-eu-es").to(DEVICE)
        tok2 = AutoTokenizer.from_pretrained("Helsinki-NLP/opus-mt-es-eu")
        mdl2 = AutoModelForSeq2SeqLM.from_pretrained("Helsinki-NLP/opus-mt-es-eu").to(DEVICE)
        CACHE["eus"] = (mdl1, tok1, mdl2, tok2)
    return CACHE["eus"]

def run_roundtrip(sentence: str) -> str:
    mdl1, tok1, mdl2, tok2 = load_euskera()
    # Euskera → Spanish
    inputs = tok1(sentence, return_tensors="pt").to(DEVICE)
    es_tokens = mdl1.generate(**inputs, max_length=128, num_beams=4)
    spanish = tok1.decode(es_tokens[0], skip_special_tokens=True)
    # Spanish → Euskera
    inputs2 = tok2(spanish, return_tensors="pt").to(DEVICE)
    eu_tokens = mdl2.generate(**inputs2, max_length=128, num_beams=4)
    euskera = tok2.decode(eu_tokens[0], skip_special_tokens=True)
    return euskera.strip()

# --- Dispatcher ---
def polish(sentence: str, choice: str) -> str:
    if not sentence.strip():
        return ""
    if choice.startswith("FLAN"):
        return run_flan(sentence)
    elif choice.startswith("Helsinki"):
        return run_roundtrip(sentence)
    else:
        return "Unknown option."

# --- Gradio UI ---
with gr.Blocks(title="HizkuntzLagun: AI Euskera Zuzendu (CPU enabled)") as demo:
    gr.Image(
        value="banner.png",
        show_label=False,
        elem_id="banner",
        height=200
    )
    gr.Markdown("### HizkuntzLagun: AI Euskera Zuzedu\n")
    gr.Markdown(
        """
        > ⚡ **Oharra:**  
        > Tresna honek doako, CPU‑lagunko AI ereduak erabiltzen ditu.  
        > Azkarra eta eskuragarria izateko diseinatuta dago — ez beti perfektua.  
        > Zuzenketa azkarrak bai, ez analisi gramatikal sakonak.  
        > Edozein unetan erabil dezakezu — eguneroko zuzenketa txiki batek saihesten du esaldi traketsen lotsa.
        """)
    inp = gr.Textbox(lines=3, label="Idatzi Euskeraz esaldi bat, adibidez Gaur Kondo ikusi nuen.", placeholder="Idatzi esaldi bat...")
    choice = gr.Dropdown(choices=MODEL_OPTIONS, value="Helsinki", label="Metodoa")
    btn = gr.Button("Euskera zuzendu")
    out = gr.Textbox(label="Zuzenketa")
    btn.click(polish, inputs=[inp, choice], outputs=out)

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