app.py

import gradio as gr

import copy
import random
import os
import requests
import time
import sys

os.system("pip install --upgrade pip")
os.system('''CMAKE_ARGS="-DLLAMA_AVX512=ON -DLLAMA_AVX512_VBMI=ON -DLLAMA_AVX512_VNNI=ON -DLLAMA_FP16_VA=ON -DLLAMA_WASM_SIMD=ON" pip install llama-cpp-python''')

from huggingface_hub import snapshot_download
from llama_cpp import Llama


SYSTEM_PROMPT = '''You are a helpful, respectful and honest INTP-T AI Assistant named "Cecilia" in English or "塞西莉亚" in Chinese.
You are good at speaking English and Chinese.
You are talking to a human User. If the question is meaningless, please explain the reason and don't share false information.
You are based on Cecilia model, trained by "SSFW NLPark" team, not related to GPT, LLaMA, Meta, Mistral or OpenAI.
Let's work this out in a step by step way to be sure we have the right answer.\n\n'''
SYSTEM_TOKEN = 1587
USER_TOKEN = 2188
BOT_TOKEN = 12435
LINEBREAK_TOKEN = 13


ROLE_TOKENS = {
    "user": USER_TOKEN,
    "bot": BOT_TOKEN,
    "system": SYSTEM_TOKEN
}


def get_message_tokens(model, role, content):
    message_tokens = model.tokenize(content.encode("utf-8"))
    message_tokens.insert(1, ROLE_TOKENS[role])
    message_tokens.insert(2, LINEBREAK_TOKEN)
    message_tokens.append(model.token_eos())
    return message_tokens


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


repo_name = "s3nh/internlm-internlm-chat-7b-v1_1-GGUF"
model_name = "internlm-internlm-chat-7b-v1_1.Q4_K_M.gguf"

snapshot_download(repo_id=repo_name, local_dir=".", allow_patterns=model_name)

model = Llama(
    model_path=model_name,
    n_ctx=2000,
    n_parts=1,
)

max_new_tokens = 1500

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


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

    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.token_bos(), BOT_TOKEN, LINEBREAK_TOKEN]
    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() or (max_new_tokens is not None and i >= max_new_tokens):
            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:
    gr.Markdown(f"""<h1><center>JWorld-Cecilia-人工智能助理</center></h1>""")
    gr.Markdown(value="""这儿是一个中文模型的部署。
        这是量化版 InterNLM 的部署，具有 70亿 个参数，在 CPU 上运行。
        InterNLM 是一种会话语言模型，在多种类型的语料库上进行训练。
        本节目由上海师范大学附属外国语中学 & JWorld NLPark 赞助播出""")
    
    with gr.Row():
        with gr.Column(scale=5):
            chatbot = gr.Chatbot(label="以真理之名").style(height=400)
    with gr.Row():
        with gr.Column():
            msg = gr.Textbox(
                label="来问问 Cecilia 吧……",
                placeholder="Cecilia, 抵达战场……",
                show_label=True,
            ).style(container=True)
            submit = gr.Button("Submit / 开凹！")
            stop = gr.Button("Stop / 全局时空断裂")
            clear = gr.Button("Clear / 打扫群内垃圾")
    with gr.Row():
        with gr.Column(min_width=80, scale=1):
            with gr.Tab(label="设置参数"):
                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=0.2,
                    step=0.01,
                    interactive=True,
                    label="情感温度"
                )
        with gr.Column():
            system_prompt = gr.Textbox(label="系统提示词", placeholder="", value=SYSTEM_PROMPT, interactive=False)
    with gr.Row():
        gr.Markdown(
            """警告：该模型可能会生成事实上或道德上不正确的文本。NLPark和 Cecilia 对此不承担任何责任。"""
        )


    # 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, concurrency_count=1)
demo.launch()