app.py

import logging
import gradio as gr
import torch
from transformers import AutoModelForSeq2SeqLM, NllbTokenizer

logging.basicConfig(
    level=logging.INFO,
    format='[%(asctime)s] {%(filename)s:%(lineno)d} %(levelname)s - %(message)s',
    datefmt='%H:%M:%S'
)
logger = logging.getLogger()

AVAILABLE_MODELS = {
    "NLLB for transliteration": "kesha-humonen/tr-eng_checkpoint-8556",
    "NLLB for hieroglyphs": "kesha-humonen/hi-eng_dpo_checkpoint-3342"
}

LANGUAGES = {
    "English": "eng_Latn",
    "German": "deu_Latn"
}

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
current_model = None
current_tokenizer = None
current_model_name = None

def unload_model():
    global current_model
    if current_model is not None:
        logger.info(f"Unloading current model: {current_model.name_or_path}")
        del current_model
        torch.cuda.empty_cache()
        current_model = None

def load_model(model_name: str):
    global current_model, current_tokenizer, current_model_name
    unload_model()
    
    logger.info(f"Loading model: {model_name}")
    
    current_model = AutoModelForSeq2SeqLM.from_pretrained(
        AVAILABLE_MODELS[model_name]
    ).to(device)
    current_tokenizer = NllbTokenizer.from_pretrained(AVAILABLE_MODELS[model_name])

    current_model_name = model_name
    return "The model has been uploaded successfully!"

def generate(input_texts: str, model_name: str, language: str) -> str:
    """Генерирует текст на основе входных данных"""
    if current_model is None or current_tokenizer is None:
        return "Please select and upload the model first."
    
    if model_name == "NLLB for transliteration":
        current_tokenizer.src_lang = 'egy_Tnt'
    elif model_name == "NLLB for hieroglyphs":
        current_tokenizer.src_lang = 'egy_Hiero'
    
    encoded_inputs = current_tokenizer(
        input_texts,
        padding=True,
        truncation=True,
        return_tensors="pt"
    ).to(current_model.device)
    
    with torch.no_grad():
        # Устанавливаем язык для вывода на основе выбора пользователя
        forced_bos_token_id = current_tokenizer.convert_tokens_to_ids(LANGUAGES[language])
        generated_tokens = current_model.generate(
            **encoded_inputs,
            forced_bos_token_id=forced_bos_token_id,
            num_beams=4,
            early_stopping=True,
            repetition_penalty=3.0
        )
    
    output_text = current_tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
    response = output_text.replace(f'{LANGUAGES[language]} ', '')
    return response

def predict(model_choice, message, language):
    global current_model_name
    if current_model is None or model_choice != current_model_name:
        load_model(model_choice)
    return generate(message, model_choice, language)

demo = gr.Interface(
    allow_flagging="never",
    fn=predict,
    inputs=[
        gr.Dropdown(
            choices=list(AVAILABLE_MODELS.keys()),
            label="Select a model",
            value=list(AVAILABLE_MODELS.keys())[0]
        ),
        gr.Textbox(
            label="Enter the sentence using transliteration or hieroglyphs.",
            placeholder="",
            lines=3
        ),
        gr.Dropdown(
            choices=list(LANGUAGES.keys()),
            label="Select output language",
            value="English"
        )
    ],
    outputs=[
        gr.Textbox(
            label="Translation",
            lines=10
        )
    ],
    title="",
    examples=[
        ["NLLB for transliteration", "wn sbte nb ẖn =f", "English"],
        ["NLLB for hieroglyphs", "𓅭 𓆑 𓉐𓉻𓌕𓏌 𓋴𓌉𓂖 𓊪𓏏𓎛𓊵𓏏𓊪", "English"],
        ["NLLB for transliteration", "m wṯs jb =k n z", "German"],
        ["NLLB for hieroglyphs", "𓍹𓅃𓇋𓂓𓅱𓍺𓌳𓍘𓃫𓌸𓊖", "German"],
    ],
    theme='base',
        #     :root {
        #     --bg: rgb(22,28,38) !important;
        #     --bg-dark: rgb(22,28,38) !important;
        #     --col: #f4f4f5 !important;
        #     --col-dark: #f4f4f5 !important;
        # }

        # body, .gradio-container, .gradio-container > div, .gradio-container .panel, .gradio-container .output, .gradio-container .input {
        #     background: rgb(22,28,38) !important;
        #     color: #f4f4f5 !important;
        #     margin: 0;
        #     padding: 0;
        #     height: 100%;
        #     width: 100%;
        #     overflow: hidden;
        # }
    css="""
        button {
            background-color: rgb(15, 138, 129) !important;
            color: white !important;
        }

        button:hover {
            background-color: rgb(15, 138, 129) !important;
        }

        footer {
            display: none !important;
        }
    """,
)

if __name__ == "__main__":
    default_model = list(AVAILABLE_MODELS.keys())[0]
    load_model(default_model)
    demo.launch(share=False)