model-ml/ml-api/app/translate.py

from functools import lru_cache
from typing import Optional

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM


BASE_MODEL_NAME = "facebook/nllb-200-distilled-600M"
CUSTOM_MODEL_DIR = "/work/models/nllb-custom"


@lru_cache(maxsize=1)
def get_tokenizer(model_name_or_path: Optional[str] = None):
    name = model_name_or_path or BASE_MODEL_NAME
    return AutoTokenizer.from_pretrained(name)


@lru_cache(maxsize=1)
def get_model(model_name_or_path: Optional[str] = None):
    name = model_name_or_path or BASE_MODEL_NAME
    return AutoModelForSeq2SeqLM.from_pretrained(name)


def load_custom_or_base():
    """Charge le modèle custom s'il existe, sinon le modèle de base."""
    import os

    if os.path.isdir(CUSTOM_MODEL_DIR):
        tok = get_tokenizer(CUSTOM_MODEL_DIR)
        mdl = get_model(CUSTOM_MODEL_DIR)
    else:
        tok = get_tokenizer(BASE_MODEL_NAME)
        mdl = get_model(BASE_MODEL_NAME)
    return tok, mdl


def translate_text(text: str, src_lang: str, tgt_lang: str) -> str:
    """
    Traduit le texte de src_lang vers tgt_lang via NLLB.
    Les codes de langue NLLB ressemblent à 'fra_Latn', 'lin_Latn', etc.
    """
    tok, mdl = load_custom_or_base()
    inputs = tok(text, return_tensors="pt")
    # Indication des langues (optionnelle, selon le tokenizer)
    # Quelques tokenizers NLLB utilisent des tokens de langue spécifiques.
    generated_tokens = mdl.generate(**inputs, max_new_tokens=200)
    out = tok.batch_decode(generated_tokens, skip_special_tokens=True)
    return out[0] if out else ""