app.py

import gradio as gr
import torch
from kuidastaltsutadalaamat.trainllm import load_model, load_tokenizer
from kuidastaltsutadalaamat.inference import llm_generate
from kuidastaltsutadalaamat.data import LazyTokenizingInferenceDataset
from kuidastaltsutadalaamat.promptops import *

accel = None
model_id = "tartuNLP/smugri4-mt"

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = load_model(model_id, device, accelerator=accel, attention="eager") #eager for cpu
model.eval()
tokenizer = load_tokenizer(model_id, accelerator=accel)
lang_raw_to_label = {"English": "English",
 "Erzya": "Erzya",
 "Estonian": "Estonian",
 "Estonian, Alutaguse, Lüg, dictionary": "Alutaguse",
 "Estonian, Hiiu, Rei, dictionary": "Hiiu",
 "Estonian, Ida, Kod, dictionary": "Ida",
 "Estonian, Kesk, Kjn, dictionary": "Keskmurre",
 "Estonian, Kihnu, dictionary": "Kihnu",
 "Estonian, Lääne, Mar, dictionary": "Lääne",
 "Estonian, Muhu, dictionary": "Muhu",
 "Estonian, Ranna, Kuu, dictionary": "Rannakeel",
 "Estonian, Saare, Khk, dictionary": "Saare",
 "Southern Estonian, Mulgi, Krk, dictionary": "Mulgi",
 "Southern Estonian, Seto, dictionary": "Seto",
 "Southern Estonian, Tartu, Nõo, dictionary": "Tartu",
 "Southern Estonian, Võro, Lei, dictionary": "Leivu",
 "Southern Estonian, Võro, Lut, dictionary": "Lutsi",
 "Southern Estonian, Võro, Sõnaq": "Võro (Sõnaq orth)",
 "Southern Estonian, Võro, Uma": "Võro (Umaleht orth)",
 "Finnish": "Finnish",
 "Kven": "Kven",
 "Meänkieli": "Meänkieli",
 "Hill Mari": "Hill Mari",
 "Meadow Mari": "Meadow Mari",
 "Hungarian": "Hungarian",
 "Inari Sami": "Inari Sami",
 "Pite Sami": "Pite Sami",
 "Kildin Sami, Antonova": "Kildin Sami (Antonova orth)",
 "Kildin Sami, Kuruch": "Kildin Sami (Kuruch orth)",
 "Lule Sami": "Lule Sami",
 "Northern Sami": "Northern Sami",
 "Skolt Sami": "Skolt Sami",
 "Southern Sami": "Southern Sami",
 "Ume Sami": "Ume Sami",
 "Izhorian, Mehmet": "Ingrian (Ala-Laukaa / simplified)",
 "Izhorian, Alamaluuga, speech": "Ingrian (Ala-Laukaa)",
 "Izhorian, Soikkola": "Ingrian (Soikkola)",
 "Votic, Standard": "Votic",
 "Komi-Permyak": "Komi-Permyak",
 "Komi-Zyrian": "Komi-Zyrian",
 "Latvian": "Latvian",
 "Livonian, Standard": "Livonian",
 "Livvi, Newwritten": "Livvi",
 "Ludian, Miikul": "Ludian (ü)",
 "Ludian, Newwritten": "Ludian (y)",
 "Mansi, Unk": "Mansi (Northern)",
 "Moksha": "Moksha",
 "Norwegian": "Norwegian",
 "Kazym Khanty, 2013": "Kazym Khanty",
 "Priur Khanty": "Priur Khanty",
 "Shur Khanty, 2013": "Shur Khanty",
 "Sred Khanty": "Sred Khanty",
 "Surgut Khanty, 2013": "Surgut Khanty",
 "Vakh Khanty, 2013": "Vakh Khanty",
 "Proper Karelian, Newwritten": "Proper Karelian",
 "Russian": "Russian",
 "Swedish": "Swedish",
 "Udmurt": "Udmurt",
 "Veps, Newwritten": "Veps"
                }

label_to_raw = { e[1]: e[0] for e in lang_raw_to_label.items() }

languages_labels = sorted(list(label_to_raw.keys()))

def run_inference(text, from_lang, to_lang, mode):
    entry = {"src_segm": text, "task": mode}
    if mode == "translate":
        entry.update({"src_lang": label_to_raw[from_lang], "tgt_lang": label_to_raw[to_lang]})
        prompt_format = PF_SMUGRI_MT
    else:
        prompt_format = PF_SMUGRI_LID

    ds = LazyTokenizingInferenceDataset([entry], tokenizer, prompt_format)
    tok = ds[0]
    output = llm_generate(model, tokenizer, tok, debug=False, max_len=512)
    return output[0]

with gr.Blocks() as demo:
    text_input = gr.Textbox(label="Text", lines=6, placeholder="Enter text...")
    #identify_btn = gr.Button("Identify language", interactive=False)
    with gr.Row():
        from_dropdown = gr.Dropdown(choices=languages_labels, label="From", value=None)
        to_dropdown = gr.Dropdown(choices=languages_labels, label="To", value=None)
    translate_btn = gr.Button("Translate", interactive=False)
    output = gr.Textbox(label="Output", lines=6)

    #def toggle_identify(text):
    #    return gr.update(interactive=bool(text.strip()))
    #text_input.change(toggle_identify, [text_input], [identify_btn])

    def toggle_translate(text, f, t):
        return gr.update(interactive=bool(text.strip() and f and t))
    text_input.change(toggle_translate, [text_input, from_dropdown, to_dropdown], [translate_btn])
    from_dropdown.change(toggle_translate, [text_input, from_dropdown, to_dropdown], [translate_btn])
    to_dropdown.change(toggle_translate, [text_input, from_dropdown, to_dropdown], [translate_btn])

    #identify_btn.click(
    #    fn=lambda text: run_inference(text, None, None, mode="identify"),
    #    inputs=[text_input],
    #    outputs=[from_dropdown],
    #).then(
    #    lambda *args: gr.update(interactive=bool(text_input.value.strip() and from_dropdown.value and to_dropdown.value)),
    #    [], [translate_btn]
    #)

    translate_btn.click(
        fn=lambda text, f, t: run_inference(text, f, t, mode="translate"),
        inputs=[text_input, from_dropdown, to_dropdown],
        outputs=[output]
    )

if __name__ == "__main__":
    demo.launch()