app.py

import gradio as gr
import logging
import tempfile
import os

# Same logger setup
logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s")
logger = logging.getLogger(__name__)

try:
    # We use "NllbTokenizer" and "AutoModelForSeq2SeqLM" from HF
    from transformers import NllbTokenizer, AutoModelForSeq2SeqLM
except ImportError:
    logger.error("transformers library not found. Ensure 'transformers' is in requirements.txt.")
    raise

# We'll assume your HF model is publicly available at "zensalaria/my-nllb-distilled"
MODEL_NAME = "zensalaria/my-nllb-distilled"

logger.info("Loading NLLB model from Hugging Face...")
try:
    tokenizer = NllbTokenizer.from_pretrained(MODEL_NAME)
    model = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
    logger.info("Model and tokenizer loaded successfully.")
except Exception as e:
    logger.error(f"Error loading model: {e}")
    raise e

def translate_text(input_text, target_lang="urd_Arab", max_length=512):
    """
    Replicates your run script's translation logic, but in-memory (no local file writes).
    """
    logger.info(f"Translating text to {target_lang}...")
    try:
        inputs = tokenizer(
            input_text, 
            return_tensors="pt", 
            max_length=max_length, 
            truncation=True
        )

        # Forced BOS token for target language
        if hasattr(tokenizer, "lang_code_to_id"):
            inputs["forced_bos_token_id"] = tokenizer.lang_code_to_id[target_lang]
        else:
            inputs["forced_bos_token_id"] = tokenizer.convert_tokens_to_ids(target_lang)

        outputs = model.generate(**inputs, max_length=max_length)
        translated = tokenizer.decode(outputs[0], skip_special_tokens=True)
        logger.info("Translation complete!")
        return translated
    except Exception as e:
        logger.error(f"Error during translation: {e}")
        return "Error translating text"

def process_translation_request(input_text, target_lang="urd_Arab"):
    """
    Logic from process_translation_request, but uses the in-memory translate_text.
    """
    if not input_text.strip():
        return "Error: No text provided."

    return translate_text(input_text, target_lang)

def gradio_interface(text, lang):
    return process_translation_request(text, lang)

# Example language list for Gradio
LANG_CHOICES = [
    ("English (Latin)", "eng_Latn"),
    ("Urdu (Arabic)", "urd_Arab"),
    ("Spanish (Latin)", "spa_Latn"),
]

demo = gr.Interface(
    fn=gradio_interface,
    inputs=[
        gr.Textbox(label="Input Text", lines=3),
        gr.Dropdown(
            choices=[c[1] for c in LANG_CHOICES],
            label="Target Language",
            value="urd_Arab"
        )
    ],
    outputs="text",
    title="NLLB-200 Translator",
    description="Translate text using your NLLB-200-distilled-600M model."
)

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