app.py

# app.py
# Requirements: transformers, torch, sentencepiece, sacremoses, gradio

import torch
import gradio as gr
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, MarianMTModel, MarianTokenizer

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

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

# Cache
CACHE = {}

# --- FLAN loader ---
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"Correct grammar and rewrite in fluent British English: {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()

# --- Marian round-trip loader ---
def load_marian():
    if "en_es" not in CACHE:
        tok1 = MarianTokenizer.from_pretrained("Helsinki-NLP/opus-mt-en-es")
        mdl1 = MarianMTModel.from_pretrained("Helsinki-NLP/opus-mt-en-es").to(DEVICE)
        tok2 = MarianTokenizer.from_pretrained("Helsinki-NLP/opus-mt-es-en")
        mdl2 = MarianMTModel.from_pretrained("Helsinki-NLP/opus-mt-es-en").to(DEVICE)
        CACHE["en_es"] = (mdl1, tok1, mdl2, tok2)
    return CACHE["en_es"]

def run_roundtrip(sentence: str) -> str:
    mdl1, tok1, mdl2, tok2 = load_marian()
    # English → 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 → English
    inputs2 = tok2(spanish, return_tensors="pt").to(DEVICE)
    en_tokens = mdl2.generate(**inputs2, max_length=128, num_beams=4)
    english = tok2.decode(en_tokens[0], skip_special_tokens=True)
    return english.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("Round-trip"):
        return run_roundtrip(sentence)
    else:
        return "Unknown option."

# --- Gradio UI ---
with gr.Blocks(title="English Grammar Polisher") as demo:
    gr.Markdown("### English Grammar Polisher\nChoose FLAN-T5 (Google) or OPUS-MT round-trip (Helsinki-NLP).")
    inp = gr.Textbox(lines=3, label="Input (English)", placeholder="Type a sentence…")
    choice = gr.Dropdown(choices=MODEL_OPTIONS, value="FLAN-T5-base (Google en→en)", label="Method")
    btn = gr.Button("Polish")
    out = gr.Textbox(label="Output")
    btn.click(polish, inputs=[inp, choice], outputs=out)

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