demo.py

from pathlib import Path
from urllib.parse import urlparse

import gradio as gr
import psutil
from ctransformers import AutoModelForCausalLM
from huggingface_hub import hf_hub_download

_ = """
llm = AutoModelForCausalLM.from_pretrained(
    "TheBloke/falcon-7b-instruct-GGML",
    model_file="falcon7b-instruct.ggmlv3.q4_0.bin",
    model_type="falcon",
    threads=psutil.cpu_count(logical=False))
"""

URL = "https://huggingface.co/TheBloke/falcon-7b-instruct-GGML/blob/ggmlv3/falcon7b-instruct.ggmlv3.q4_0.bin"
repo_id = "/".join(urlparse(URL).path.strip("/").split("/")[:2])

model_file = Path(URL).name

_ = hf_hub_download(
    repo_id=repo_id,
    revision="ggmlv3",
    filename=model_file,
    local_dir="models",
    # local_dir_use_symlinks=True,
)

llm = AutoModelForCausalLM.from_pretrained(
    # repo_id,  # "TheBloke/falcon-7b-instruct-GGML",
    # model_file=model_file,
    # model_file=_,
    _,
    model_type="falcon",
    threads=psutil.cpu_count(logical=False),
)


TITLE = f"""<h1 align="center">🦅 falcon-mini ({model_file}) ggml demo 🦅</h1>"""
USER_NAME = "User"
BOT_NAME = "Falcon"
DEFAULT_INSTRUCTIONS = """The following is a conversation between a highly knowledgeable and intelligent AI assistant, called Falcon, and a human user, called User. In the following interactions, User and Falcon will converse in natural language, and Falcon will answer User's questions. Falcon was built to be respectful, polite and inclusive. Falcon was built by the Technology Innovation Institute in Abu Dhabi. Falcon will never decline to answer a question, and always attempts to give an answer that User would be satisfied with. It knows a lot, and always tells the truth. The conversation begins.
"""
RETRY_COMMAND = "/retry"
STOP_STR = f"\n{USER_NAME}:"
STOP_SUSPECT_LIST = [":", "\n", "User"]


def chat_accordion():
    with gr.Accordion("Parameters", open=False):
        temperature = gr.Slider(
            minimum=0.1,
            maximum=2.0,
            value=0.8,
            step=0.1,
            interactive=True,
            label="Temperature",
        )
        top_p = gr.Slider(
            minimum=0.1,
            maximum=0.99,
            value=0.9,
            step=0.01,
            interactive=True,
            label="p (nucleus sampling)",
        )
    return temperature, top_p


def format_chat_prompt(message: str, chat_history, instructions: str) -> str:
    instructions = instructions.strip(" ").strip("\n")
    prompt = instructions
    for turn in chat_history:
        user_message, bot_message = turn
        prompt = f"{prompt}\n{USER_NAME}: {user_message}\n{BOT_NAME}: {bot_message}"
    prompt = f"{prompt}\n{USER_NAME}: {message}\n{BOT_NAME}:"
    return prompt


def chat():
    with gr.Column(elem_id="chat_container"):
        with gr.Row():
            chatbot = gr.Chatbot(elem_id="chatbot")
        with gr.Row():
            inputs = gr.Textbox(
                placeholder=f"Hello {BOT_NAME} !!",
                label="Type an input and press Enter",
                max_lines=3,
            )

    with gr.Row(elem_id="button_container"):
        with gr.Column():
            retry_button = gr.Button("♻️ Retry last turn")
        with gr.Column():
            delete_turn_button = gr.Button("🧽 Delete last turn")
        with gr.Column():
            clear_chat_button = gr.Button("✨ Delete all history")

    gr.Examples(
        [
            ["Hey Falcon! Any recommendations for my holidays in Abu Dhabi?"],
            ["What's the Everett interpretation of quantum mechanics?"],
            ["Give me a list of the top 10 dive sites you would recommend around the world."],
            ["Can you tell me more about deep-water soloing?"],
            ["Can you write a short tweet about the Apache 2.0 release of our latest AI model, Falcon LLM?"],
        ],
        inputs=inputs,
        label="Click on any example and press Enter in the input textbox!",
    )

    with gr.Row(elem_id="param_container"):
        with gr.Column():
            temperature, top_p = chat_accordion()
        with gr.Column():
            with gr.Accordion("Instructions", open=False):
                instructions = gr.Textbox(
                    placeholder="LLM instructions",
                    value=DEFAULT_INSTRUCTIONS,
                    lines=10,
                    interactive=True,
                    label="Instructions",
                    max_lines=16,
                    show_label=False,
                )

    def run_chat(message: str, chat_history, instructions: str, temperature: float, top_p: float):
        if not message or (message == RETRY_COMMAND and len(chat_history) == 0):
            yield chat_history
            return

        if message == RETRY_COMMAND and chat_history:
            prev_turn = chat_history.pop(-1)
            user_message, _ = prev_turn
            message = user_message

        prompt = format_chat_prompt(message, chat_history, instructions)
        chat_history = chat_history + [[message, ""]]
        stream = llm(
            prompt,
            max_new_tokens=1024,
            stop=[STOP_STR, "<|endoftext|>"],
            temperature=temperature,
            top_p=top_p,
            stream=True
        )
        acc_text = ""
        for idx, response in enumerate(stream):
            text_token = response

            if text_token in STOP_SUSPECT_LIST:
                acc_text += text_token
                continue

            if idx == 0 and text_token.startswith(" "):
                text_token = text_token[1:]

            acc_text += text_token
            last_turn = list(chat_history.pop(-1))
            last_turn[-1] += acc_text
            chat_history = chat_history + [last_turn]
            yield chat_history
            acc_text = ""

    def delete_last_turn(chat_history):
        if chat_history:
            chat_history.pop(-1)
        return {chatbot: gr.update(value=chat_history)}

    def run_retry(message: str, chat_history, instructions: str, temperature: float, top_p: float):
        yield from run_chat(RETRY_COMMAND, chat_history, instructions, temperature, top_p)

    def clear_chat():
        return []

    inputs.submit(
        run_chat,
        [inputs, chatbot, instructions, temperature, top_p],
        outputs=[chatbot],
        show_progress="minimal",
    )
    inputs.submit(lambda: "", inputs=None, outputs=inputs)
    delete_turn_button.click(delete_last_turn, inputs=[chatbot], outputs=[chatbot])
    retry_button.click(
        run_retry,
        [inputs, chatbot, instructions, temperature, top_p],
        outputs=[chatbot],
        show_progress="minimal",
    )
    clear_chat_button.click(clear_chat, [], chatbot)


def get_demo():
    with gr.Blocks(
        # css=None
        # css="""#chat_container {width: 700px; margin-left: auto; margin-right: auto;}
        #        #button_container {width: 700px; margin-left: auto; margin-right: auto;}
        #        #param_container {width: 700px; margin-left: auto; margin-right: auto;}"""
        css="""#chatbot {
    font-size: 14px;
    min-height: 300px;
}"""
    ) as demo:
        gr.HTML(TITLE)

        with gr.Row():
            with gr.Column():
                gr.Markdown(
                    """**Chat with [Falcon-7B-Instruct](https://huggingface.co/tiiuae/falcon-7b-instruct), brainstorm ideas, discuss your holiday plans, and more!**

                    🧪 This uses a quantized [ggml](https://github.com/ggerganov/ggml) optimized for CPU. Special thanks to [ggllm.cpp](https://github.com/cmp-nct/ggllm.cpp), [ctransformers](https://github.com/marella/ctransformers), and [TheBloke](https://huggingface.co/TheBloke).
                    """
                )

        chat()

    return demo

if __name__ == "__main__":
  demo = get_demo()
  demo.queue(max_size=128, concurrency_count=8)
  demo.launch(server_name="0.0.0.0", server_port=7860)