en_dv_latin.py

import random
import numpy as np
import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import spaces


# Available models
MODEL_OPTIONS_TRANSLATE = {
    #"Flan-T5-A-Dhivehi-Latin Model": "alakxender/flan-t5-base-dhivehi-en-latin-v2",
    "Flan-T5-B-Dhivehi-Latin Model": "alakxender/flan-t5-base-dhivehi-en-latin",
    "MT5-B-Dhivehi-English Model": "alakxender/mt5-base-dv-en",
    "MT5-B1-Dhivehi-English Model": "alakxender/mt5-base-dv-en-md",
}

# Cache for loaded models/tokenizers
MODEL_CACHE = {}

def get_model_and_tokenizer(model_dir):
    if model_dir not in MODEL_CACHE:
        print(f"Loading model: {model_dir}")
        tokenizer = AutoTokenizer.from_pretrained(model_dir)
        model = AutoModelForSeq2SeqLM.from_pretrained(model_dir)
        device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        print(f"Moving model to device: {device}")
        model.to(device)
        MODEL_CACHE[model_dir] = (tokenizer, model)
    return MODEL_CACHE[model_dir]

max_input_length = 128
max_output_length = 128

@spaces.GPU()
def translate(instruction, input_text, model_choice, max_new_tokens=128, num_beams=4, repetition_penalty=1.2, no_repeat_ngram_size=3):
    model_dir = MODEL_OPTIONS_TRANSLATE[model_choice]
    tokenizer, model = get_model_and_tokenizer(model_dir)

    combined_input = f"{instruction.strip()} {input_text.strip()}" if input_text else instruction.strip()
    inputs = tokenizer(
        combined_input,
        return_tensors="pt",
        truncation=True,
        max_length=max_input_length
    )
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    inputs = {k: v.to(device) for k, v in inputs.items()}

    gen_kwargs = {
        **inputs,
        "max_length":max_new_tokens,
        "min_length":10,
        "num_beams":num_beams,
        "early_stopping":True,
        "no_repeat_ngram_size":no_repeat_ngram_size,
        "repetition_penalty":repetition_penalty,
        "do_sample":False,
        "pad_token_id":tokenizer.pad_token_id,
        "eos_token_id":tokenizer.eos_token_id
    }


    with torch.no_grad():
        outputs = model.generate(**gen_kwargs)
    decoded_output = tokenizer.decode(outputs[0], skip_special_tokens=True)

    return decoded_output