goai_helpers/goai_traduction.py

import torch
import spaces
from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig
from peft import PeftModel, PeftConfig
import os
import unicodedata
from huggingface_hub import login

max_length = 512
auth_token = os.getenv('HF_SPACE_TOKEN')
login(token=auth_token)


@spaces.GPU
def goai_traduction(text, src_lang, tgt_lang):
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    
    if src_lang == "fra_Latn" and tgt_lang == "mos_Latn":
        model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
    elif src_lang == "mos_Latn" and tgt_lang == "fra_Latn":
        model_id = "ArissBandoss/3b-new-400"  
    else:
        model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
    
    tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_id, token=auth_token).to(device)

    print(tokenizer.model_max_length)
    print(model.model.encoder.embed_positions.weight.shape)
    # Configuration du tokenizer
    tokenizer.src_lang = src_lang
    
    # Tokenisation
    inputs = tokenizer(text, return_tensors="pt", truncation=False).to(device)
    
    # ID du token de langue cible
    tgt_lang_id = tokenizer.convert_tokens_to_ids(tgt_lang)
    
    # Génération avec les paramètres optimaux
    outputs = model.generate(
        **inputs,
        forced_bos_token_id=tgt_lang_id,
        max_new_tokens=1024,
        early_stopping=False,
        num_beams=5,
        no_repeat_ngram_size=0,
        length_penalty=1.0
    )
    
    # Décodage
    translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    
    return translation

def real_time_traduction(input_text, src_lang, tgt_lang):
    return goai_traduction(input_text, src_lang, tgt_lang)