app.py

import gradio as gr

import os

from huggingface_hub.file_download import http_get
from llama_cpp import Llama


SYSTEM_PROMPT = "Ты — Сайга, русскоязычный автоматический ассистент. Ты разговариваешь с людьми и помогаешь им."


def get_message_tokens(model, role, content):
    content = f"{role}\n{content}\n</s>"
    content = content.encode("utf-8")
    return model.tokenize(content, special=True)


def get_system_tokens(model):
    system_message = {"role": "system", "content": SYSTEM_PROMPT}
    return get_message_tokens(model, **system_message)


def load_model(
    directory: str = ".",
    model_name: str = "model-q4_K.gguf",
    model_url: str = "https://huggingface.co/IlyaGusev/saiga2_13b_gguf/resolve/main/model-q4_K.gguf"
):
    final_model_path = os.path.join(directory, model_name)
    
    print("Downloading all files...")
    if not os.path.exists(final_model_path):
        with open(final_model_path, "wb") as f:
            http_get(model_url, f)
    os.chmod(final_model_path, 0o777)
    print("Files downloaded!")
    
    model = Llama(
        model_path=final_model_path,
        n_ctx=1024
    )
    
    print("Model loaded!")
    return model


MODEL = load_model()


def user(message, history):
    new_history = history + [[message, None]]
    return "", new_history


def bot(
    history,
    system_prompt,
    top_p,
    top_k,
    temp
):
    model = MODEL
    tokens = get_system_tokens(model)[:]

    for user_message, bot_message in history[:-1]:
        message_tokens = get_message_tokens(model=model, role="user", content=user_message)
        tokens.extend(message_tokens)
        if bot_message:
            message_tokens = get_message_tokens(model=model, role="bot", content=bot_message)
            tokens.extend(message_tokens)

    last_user_message = history[-1][0]
    message_tokens = get_message_tokens(model=model, role="user", content=last_user_message)
    tokens.extend(message_tokens)

    role_tokens = model.tokenize("bot\n".encode("utf-8"), special=True)
    tokens.extend(role_tokens)
    generator = model.generate(
        tokens,
        top_k=top_k,
        top_p=top_p,
        temp=temp
    )

    partial_text = ""
    for i, token in enumerate(generator):
        if token == model.token_eos():
            break
        partial_text += model.detokenize([token]).decode("utf-8", "ignore")
        history[-1][1] = partial_text
        yield history


with gr.Blocks(
    theme=gr.themes.Soft()
) as demo:
    favicon = '<img src="https://cdn.midjourney.com/b88e5beb-6324-4820-8504-a1a37a9ba36d/0_1.png" width="48px" style="display: inline">'
    gr.Markdown(
        f"""<h1><center>{favicon}Saiga2 13B GGUF Q4_K</center></h1>

        This is a demo of a **Russian**-speaking LLaMA2-based model. If you are interested in other languages, please check other models, such as [MPT-7B-Chat](https://huggingface.co/spaces/mosaicml/mpt-7b-chat).

        Это демонстрационная версия [квантованной Сайги-2 с 13 миллиардами параметров](https://huggingface.co/IlyaGusev/saiga2_13b_ggml), работающая на CPU.

        Сайга-2 — это разговорная языковая модель, которая основана на [LLaMA-2](https://ai.meta.com/llama/) и дообучена на корпусах, сгенерированных ChatGPT, таких как [ru_turbo_alpaca](https://huggingface.co/datasets/IlyaGusev/ru_turbo_alpaca), [ru_turbo_saiga](https://huggingface.co/datasets/IlyaGusev/ru_turbo_saiga) и [gpt_roleplay_realm](https://huggingface.co/datasets/IlyaGusev/gpt_roleplay_realm).
        """
    )
    with gr.Row():
        with gr.Column(scale=5):
            system_prompt = gr.Textbox(label="Системный промпт", placeholder="", value=SYSTEM_PROMPT, interactive=False)
            chatbot = gr.Chatbot(label="Диалог")
        with gr.Column(min_width=80, scale=1):
            with gr.Tab(label="Параметры генерации"):
                top_p = gr.Slider(
                    minimum=0.0,
                    maximum=1.0,
                    value=0.9,
                    step=0.05,
                    interactive=True,
                    label="Top-p",
                )
                top_k = gr.Slider(
                    minimum=10,
                    maximum=100,
                    value=30,
                    step=5,
                    interactive=True,
                    label="Top-k",
                )
                temp = gr.Slider(
                    minimum=0.0,
                    maximum=2.0,
                    value=0.01,
                    step=0.01,
                    interactive=True,
                    label="Температура"
                )
    with gr.Row():
        with gr.Column():
            msg = gr.Textbox(
                label="Отправить сообщение",
                placeholder="Отправить сообщение",
                show_label=False,
            )
        with gr.Column():
            with gr.Row():
                submit = gr.Button("Отправить")
                stop = gr.Button("Остановить")
                clear = gr.Button("Очистить")
    with gr.Row():
        gr.Markdown(
            """ПРЕДУПРЕЖДЕНИЕ: Модель может генерировать фактически или этически некорректные тексты. Мы не несём за это ответственность."""
        )

    # Pressing Enter
    submit_event = msg.submit(
        fn=user,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=False,
    ).success(
        fn=bot,
        inputs=[
            chatbot,
            system_prompt,
            top_p,
            top_k,
            temp
        ],
        outputs=chatbot,
        queue=True,
    )

    # Pressing the button
    submit_click_event = submit.click(
        fn=user,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=False,
    ).success(
        fn=bot,
        inputs=[
            chatbot,
            system_prompt,
            top_p,
            top_k,
            temp
        ],
        outputs=chatbot,
        queue=True,
    )

    # Stop generation
    stop.click(
        fn=None,
        inputs=None,
        outputs=None,
        cancels=[submit_event, submit_click_event],
        queue=False,
    )

    # Clear history
    clear.click(lambda: None, None, chatbot, queue=False)

demo.queue(max_size=128)
demo.launch(show_error=True)