src/runtime/translator.py

import torch
from optimum.onnxruntime import ORTModelForSeq2SeqLM
from transformers import AutoTokenizer
import os

class OfflineTranslator:
    def __init__(self, model_id="Xenova/opus-mt-es-en"):
        self.model_id = model_id
        self.tokenizer = None
        self.model = None

    def _init_model(self):
        if not self.tokenizer:
            self.tokenizer = AutoTokenizer.from_pretrained(self.model_id)
        if not self.model:
            self.model = ORTModelForSeq2SeqLM.from_pretrained(self.model_id)

    def translate(self, text: str) -> str:
        if not text or not any(c.isalpha() for c in text):
            return text
            
        self._init_model()
        
        # We always translate if the translator is active for the session.
        # The Opus-MT model is smart enough to leave existing English words as-is.
        inputs = self.tokenizer(text, return_tensors="pt", padding=True, truncation=True)
        outputs = self.model.generate(**inputs)
        translated = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        return translated