app.py

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, TextIteratorStreamer
from transformers.generation import LogitsProcessor
from threading import Thread
import gradio as gr


print(f"Starting to load the model to memory")

tokenizer = AutoTokenizer.from_pretrained("nort5_en-no_base")
cls_index = tokenizer.convert_tokens_to_ids("[CLS]")
sep_index = tokenizer.convert_tokens_to_ids("[SEP]")
eos_index = tokenizer.convert_tokens_to_ids("[EOS]")
pad_index = tokenizer.convert_tokens_to_ids("[PAD]")
eng_index = tokenizer.convert_tokens_to_ids(">>eng<<")
nob_index = tokenizer.convert_tokens_to_ids(">>nob<<")
nno_index = tokenizer.convert_tokens_to_ids(">>nno<<")

model = AutoModelForSeq2SeqLM.from_pretrained("nort5_en-no_base", trust_remote_code=True)

device = "cuda" if torch.cuda.is_available() else "cpu"
print(f"SYSTEM: Running on {device}", flush=True)

model = model.to(device)
model.eval()

print(f"Sucessfully loaded the model to the memory")


LANGUAGES = [
    "🇬🇧 English",
    "🇳🇴 Norwegian (Bokmål)",
    "🇳🇴 Norwegian (Nynorsk)"
]

LANGUAGE_IDS = {
    "🇬🇧 English": eng_index,
    "🇳🇴 Norwegian (Bokmål)": nob_index,
    "🇳🇴 Norwegian (Nynorsk)": nno_index
}


class BatchStreamer(TextIteratorStreamer):
    def put(self, value):
        print(value.shape)

        #if value.size(0) == 1:
        #    return super().put(value)

        if len(self.token_cache) == 0:
            self.token_cache = [[] for _ in range(value.size(0))]

        value = value.tolist()

        # Add the new token to the cache and decodes the entire thing.
        for c, v in zip(self.token_cache, value):
            c += [v] if isinstance(v, int) else v

        paragraphs = [tokenizer.decode(c, **self.decode_kwargs).strip() for c in self.token_cache]
        text = '\n'.join(paragraphs)

        self.on_finalized_text(text)

    def end(self):
        if len(self.token_cache) > 0:
            paragraphs = [tokenizer.decode(c, **self.decode_kwargs).strip() for c in self.token_cache]
            printable_text = '\n'.join(paragraphs)
            self.token_cache = []
            self.print_len = 0
        else:
            printable_text = ""

        self.next_tokens_are_prompt = True
        self.on_finalized_text(printable_text, stream_end=True)


class RepetitionPenaltyLogitsProcessor(LogitsProcessor):
    def __init__(self, penalty: float, model):
        last_bias = model.classifier.nonlinearity[-1].bias.data
        last_bias = torch.nn.functional.log_softmax(last_bias)
        self.penalty = penalty * (last_bias - last_bias.max())

    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
        penalized_score = torch.gather(scores + self.penalty.unsqueeze(0).to(input_ids.device), 1, input_ids)
        scores.scatter_(1, input_ids, penalized_score)
        return scores


def translate(source, source_language, target_language):
    if source_language == target_language:
        yield source.strip()
        return source.strip()

    source = [s.strip() for s in source.split('\n')]
    source_subwords = tokenizer(source).input_ids
    source_subwords = [[cls_index, LANGUAGE_IDS[target_language], LANGUAGE_IDS[source_language]] + s + [sep_index] for s in source_subwords]
    source_subwords = [torch.tensor(s) for s in source_subwords]
    source_subwords = torch.nn.utils.rnn.pad_sequence(source_subwords, batch_first=True, padding_value=pad_index)
    source_subwords = source_subwords[:, :512].to(device)

    streamer = BatchStreamer(tokenizer, timeout=60.0, skip_special_tokens=True)

    def generate(model, **kwargs):
        with torch.inference_mode():
            with torch.autocast(enabled=device != "cpu", device_type=device, dtype=torch.bfloat16):
                return model.generate(**kwargs)

    generate_kwargs = dict(
        streamer=streamer,
        input_ids=source_subwords,
        attention_mask=(source_subwords != pad_index).long(),
        max_new_tokens = 512-1,
        #top_k=64,
        #top_p=0.95,
        #do_sample=True,
        #temperature=0.3,
        num_beams=1,
        #use_cache=True,
        logits_processor=[RepetitionPenaltyLogitsProcessor(1.0, model)],
        # num_beams=4,
        # early_stopping=True,
        do_sample=False,
        use_cache=True
    )
    t = Thread(target=generate, args=(model,), kwargs=generate_kwargs)
    t.start()

    for new_text in streamer:
        yield new_text.strip()
    return new_text.strip()


def switch_inputs(source, target, source_language, target_language):
    return target, source, target_language, source_language


with gr.Blocks(theme='sudeepshouche/minimalist') as demo:

    gr.Markdown("# Norwegian-English translation")

    with gr.Row():
        with gr.Column(scale=7, variant="panel"):
            source_language = gr.Dropdown(
                LANGUAGES, value=LANGUAGES[1], show_label=False
            )
            source = gr.Textbox(
                label="Source text", placeholder="What do you want to translate?", show_label=False, lines=7, max_lines=100, autofocus=True
            )  # .style(container=False)
            submit = gr.Button("Submit", variant="primary")  # .style(full_width=True)

        with gr.Column(scale=7, variant="panel"):
            target_language = gr.Dropdown(
                LANGUAGES, value=LANGUAGES[0], show_label=False
            )
            target = gr.Textbox(
                label="Translation", show_label=False, interactive=False, lines=7, max_lines=100
            )


    def update_state_after_user():
        return {
            source: gr.update(interactive=False),
            submit: gr.update(interactive=False),
            source_language: gr.update(interactive=False),
            target_language: gr.update(interactive=False)
        }

    def update_state_after_return():
        return {
            source: gr.update(interactive=True),
            submit: gr.update(interactive=True),
            source_language: gr.update(interactive=True),
            target_language: gr.update(interactive=True)
        }


    submit_event = source.submit(
        fn=update_state_after_user, inputs=None, outputs=[source, submit, source_language, target_language], queue=False
    ).then(
        fn=translate, inputs=[source, source_language, target_language], outputs=[target], queue=True
    ).then(
        fn=update_state_after_return, inputs=None, outputs=[source, submit, source_language, target_language], queue=False
    )
    
    submit_click_event = submit.click(
        fn=update_state_after_user, inputs=None, outputs=[source, submit, source_language, target_language], queue=False
    ).then(
        fn=translate, inputs=[source, source_language, target_language], outputs=[target], queue=True
    ).then(
        fn=update_state_after_return, inputs=None, outputs=[source, submit, source_language, target_language], queue=False
    )

demo.queue(max_size=32, concurrency_count=2)
demo.launch()