import torch
import spaces
from transformers import pipeline, AutoModelForSeq2SeqLM, AutoTokenizer
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/mos2fr-3B-1200"
    else:
        model_id = "ArissBandoss/nllb-200-distilled-600M-finetuned-fr-to-mos-V4"
    
    tokenizer = AutoTokenizer.from_pretrained(model_id, token=auth_token, truncation=True, max_length=512)
    model = AutoModelForSeq2SeqLM.from_pretrained(model_id, token=auth_token).to(device)
    
    # Ajout du code de langue source
    tokenizer.src_lang = src_lang
    
    # Tokenisation du texte d'entrée
    inputs = tokenizer(text, return_tensors="pt").to(device)
    print(inputs)
    
    # Utilisation de convert_tokens_to_ids au lieu de lang_code_to_id
    tgt_lang_id = tokenizer.convert_tokens_to_ids(tgt_lang)
    
    # Génération avec paramètres améliorés
    outputs = model.generate(
        **inputs,
        forced_bos_token_id=tgt_lang_id,
        eos_token_id=tokenizer.eos_token_id
        max_length=512,
           num_beams=4,
    do_sample=False,
    no_repeat_ngram_size=3
    )

    print("Token IDs:", outputs)
    print("Tokens:", [tokenizer.decode([tok]) for tok in outputs[0]])

    
    # Décodage de la sortie
    translation = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
    print("ici translation")
    print(translation)
    return translation


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