app.py

"""Run codes"""
# import gradio

# gradio.load("models/WizardLM/WizardCoder-15B-V1.0").launch()
import os
import time
from types import SimpleNamespace

import gradio as gr
from about_time import about_time
from ctransformers import AutoConfig, AutoModelForCausalLM
from huggingface_hub import hf_hub_download
from loguru import logger

os.environ["TZ"] = "Asia/Shanghai"
try:
    time.tzset()  # type: ignore # pylint: disable=no-member
except Exception:
    # Windows
    logger.warning("Windows, cant run time.tzset()")

ns = SimpleNamespace(
    response="",
    generator=[],
)


def predict(prompt, bot):
    # logger.debug(f"{prompt=}, {bot=}, {timeout=}")
    logger.debug(f"{prompt=}, {bot=}")

    ns.response = ""
    with about_time() as atime:  # type: ignore
        try:
            # user_prompt = prompt
            generator = generate(llm, generation_config, system_prompt, prompt.strip())
            print(assistant_prefix, end=" ", flush=True)

            response = ""
            buff.update(value="diggin...")

            for word in generator:
                # print(word, end="", flush=True)
                print(word, flush=True)  # vertical stream
                response += word
                ns.response = response
                buff.update(value=response)
            print("")
            logger.debug(f"{response=}")
        except Exception as exc:
            logger.error(exc)
            response = f"{exc=}"

    # bot = {"inputs": [response]}
    _ = (
        f"(time elapsed: {atime.duration_human}, "  # type: ignore
        f"{atime.duration/(len(prompt) + len(response)):.1f}s/char)"  # type: ignore
    )

    bot.append([prompt, f"{response} {_}"])

    return prompt, bot


def predict_api(prompt):
    logger.debug(f"{prompt=}")
    ns.response = ""
    try:
        # user_prompt = prompt
        generator = generate(llm, generation_config, system_prompt, prompt.strip())
        print(assistant_prefix, end=" ", flush=True)

        response = ""
        buff.update(value="diggin...")
        for word in generator:
            print(word, end="", flush=True)
            response += word
            ns.response = response
            buff.update(value=response)
        print("")
        logger.debug(f"{response=}")
    except Exception as exc:
        logger.error(exc)
        response = f"{exc=}"
    # bot = {"inputs": [response]}
    # bot = [(prompt, response)]

    return response


def download_quant(destination_folder: str, repo_id: str, model_filename: str):
    local_path = os.path.abspath(destination_folder)
    return hf_hub_download(
        repo_id=repo_id,
        filename=model_filename,
        local_dir=local_path,
        local_dir_use_symlinks=True,
    )


logger.info("start dl")
_ = """full url: https://huggingface.co/TheBloke/mpt-30B-chat-GGML/blob/main/mpt-30b-chat.ggmlv0.q4_1.bin"""

# https://huggingface.co/TheBloke/mpt-30B-chat-GGML
_ = """
mpt-30b-chat.ggmlv0.q4_0.bin 	q4_0 	4 	16.85 GB 	19.35 GB 	4-bit.
mpt-30b-chat.ggmlv0.q4_1.bin 	q4_1 	4 	18.73 GB 	21.23 GB 	4-bit. Higher accuracy than q4_0 but not as high as q5_0. However has quicker inference than q5 models.
mpt-30b-chat.ggmlv0.q5_0.bin 	q5_0 	5 	20.60 GB 	23.10 GB
mpt-30b-chat.ggmlv0.q5_1.bin 	q5_1 	5 	22.47 GB 	24.97 GB
mpt-30b-chat.ggmlv0.q8_0.bin 	q8_0 	8 	31.83 GB 	34.33 GB
"""
model_filename = "mpt-30b-chat.ggmlv0.q4_1.bin"
model_filename = "WizardCoder-15B-1.0.ggmlv3.q4_0.bin"  # 10.7G
model_filename = "WizardCoder-15B-1.0.ggmlv3.q4_1.bin"  # 11.9G
destination_folder = "models"

repo_id = "TheBloke/mpt-30B-chat-GGML"
if "WizardCoder" in model_filename:
    repo_id = "TheBloke/WizardCoder-15B-1.0-GGML"
    
download_quant(destination_folder, repo_id, model_filename)

logger.info("done dl")

if "mpt" in model_filename:
    config = AutoConfig.from_pretrained("mosaicml/mpt-30b-chat", context_length=8192)
    llm = AutoModelForCausalLM.from_pretrained(
        os.path.abspath(f"models/{model_filename}"),
        model_type="mpt",
        config=config,
    )

# https://huggingface.co/spaces/matthoffner/wizardcoder-ggml/blob/main/main.py
if "WizardCoder" in model_filename:
    llm = AutoModelForCausalLM.from_pretrained("TheBloke/WizardCoder-15B-1.0-GGML",
                                            model_file="",
                                            model_type="starcoder",
                                            threads=8)

system_prompt = "A conversation between a user and an LLM-based AI assistant named Local Assistant. Local Assistant gives helpful and honest answers."

user_prefix = "[user]: "
assistant_prefix = "[assistant]: "

css = """
    .importantButton {
        background: linear-gradient(45deg, #7e0570,#5d1c99, #6e00ff) !important;
        border: none !important;
    }
    .importantButton:hover {
        background: linear-gradient(45deg, #ff00e0,#8500ff, #6e00ff) !important;
        border: none !important;
    }
    .disclaimer {font-variant-caps: all-small-caps; font-size: xx-small;}
    .xsmall {font-size: x-small;}
"""

with gr.Blocks(
    # title="mpt-30b-chat-ggml",
    title=f"{model_filename}",
    theme=gr.themes.Soft(text_size="sm", spacing_size="sm"),
    css=css,
) as block:
    with gr.Accordion("🎈 Info", open=False):
        # gr.HTML(
        #     """<center><a href="https://huggingface.co/spaces/mikeee/mpt-30b-chat?duplicate=true"><img src="https://bit.ly/3gLdBN6" alt="Duplicate"></a> and spin a CPU UPGRADE to avoid the queue</center>"""
        # )
        gr.Markdown(
            f"""<h4><center>{model_filename}</center></h4>

            The examples are meant for another model. You should try with
            some coder-related prompts.

            Try to refresh the browser and try again when occasionally errors occur.

            It takes about >100 seconds to get a response. Restarting the space takes about 5 minutes if the space is asleep due to inactivity. If the space crashes for some reason, it will also take about 5 minutes to restart. You need to refresh the browser to reload the new space.
            """,
            elem_classes="xsmall",
        )
    conversation = Chat()
    chatbot = gr.Chatbot(scroll_to_output=True).style(height=700)  # 500
    buff = gr.Textbox(show_label=False)
    with gr.Row():
        with gr.Column(scale=1):
            msg = gr.Textbox(
                label="Chat Message Box",
                placeholder="Ask me anything (press Enter or click Submit to send)",
                show_label=False,
            ).style(container=False)
        with gr.Column(scale=0.1):
            with gr.Row():
                submit = gr.Button("Submit", elem_classes="xsmall")
                stop = gr.Button("Stop", visible=False)
                clear = gr.Button("Clear History", visible=True)
    with gr.Row(visible=False):
        with gr.Accordion("Advanced Options:", open=False):
            with gr.Row():
                with gr.Column(scale=2):
                    system = gr.Textbox(
                        label="System Prompt",
                        value=Chat.default_system_prompt,
                        show_label=False,
                    ).style(container=False)
                with gr.Column():
                    with gr.Row():
                        change = gr.Button("Change System Prompt")
                        reset = gr.Button("Reset System Prompt")

    with gr.Accordion("Example inputs", open=True):
        etext = """In America, where cars are an important part of the national psyche, a decade ago people had suddenly started to drive less, which had not happened since the oil shocks of the 1970s. """
        examples = gr.Examples(
            examples=[
                ["Explain the plot of Cinderella in a sentence."],
                [
                    "How long does it take to become proficient in French, and what are the best methods for retaining information?"
                ],
                ["What are some common mistakes to avoid when writing code?"],
                ["Build a prompt to generate a beautiful portrait of a horse"],
                ["Suggest four metaphors to describe the benefits of AI"],
                ["Write a pop song about leaving home for the sandy beaches."],
                ["Write a summary demonstrating my ability to tame lions"],
                ["鲁迅和周树人什么关系 说中文"],
                ["鲁迅和周树人什么关系"],
                ["鲁迅和周树人什么关系 用英文回答"],
                ["从前有一头牛，这头牛后面有什么？"],
                ["正无穷大加一大于正无穷大吗？"],
                ["正无穷大加正无穷大大于正无穷大吗？"],
                ["-2的平方根等于什么"],
                ["树上有5只鸟，猎人开枪打死了一只。树上还有几只鸟？"],
                ["树上有11只鸟，猎人开枪打死了一只。树上还有几只鸟？提示：需考虑鸟可能受惊吓飞走。"],
                ["以红楼梦的行文风格写一张委婉的请假条。不少于320字。"],
                [f"{etext} 翻成中文，列出3个版本"],
                [f"{etext} \n 翻成中文，保留原意，但使用文学性的语言。不要写解释。列出3个版本"],
                ["js 判断一个数是不是质数"],
                ["js 实现python 的 range(10)"],
                ["js 实现python 的 [*(range(10)]"],
                ["假定 1 + 2 = 4, 试求 7 + 8"],
                ["Erkläre die Handlung von Cinderella in einem Satz."],
                ["Erkläre die Handlung von Cinderella in einem Satz. Auf Deutsch"],
            ],
            inputs=[msg],
            examples_per_page=40,
        )

    # with gr.Row():
    with gr.Accordion("Disclaimer", open=False):
        gr.Markdown(
            f"Disclaimer: {"-".join(model_filename.split("-")[:2])} can produce factually incorrect output, and should not be relied on to produce "
            "factually accurate information. MPT-30B was trained on various public datasets; while great efforts "
            "have been taken to clean the pretraining data, it is possible that this model could generate lewd, "
            "biased, or otherwise offensive outputs.",
            elem_classes=["disclaimer"],
        )

    msg.submit(
        # fn=conversation.user_turn,
        fn=predict0,
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
        api_name="predict",
    )
    submit.click(
        # fn=predict0,
        fn=lambda x, y: ("",) + predict0(x, y)[1:],  # clear msg
        inputs=[msg, chatbot],
        outputs=[msg, chatbot],
        queue=True,
        show_progress="full",
    )
    clear.click(lambda: None, None, chatbot, queue=False)

    # update buff Textbox, every: units in seconds)
    # https://huggingface.co/spaces/julien-c/nvidia-smi/discussions
    # does not work
    # AttributeError: 'Blocks' object has no attribute 'run_forever'
    # block.run_forever(lambda: ns.response, None, [buff], every=1)

    with gr.Accordion("For Chat/Translation API", open=False, visible=False):
        input_text = gr.Text()
        api_btn = gr.Button("Go", variant="primary")
        out_text = gr.Text()
    api_btn.click(
        predict_api,
        input_text,
        out_text,
        # show_progress="full",
        api_name="api",
    )

# concurrency_count=5, max_size=20
# max_size=36, concurrency_count=14
block.queue(concurrency_count=5, max_size=20).launch(debug=True)