app.py

from fix_int8 import fix_pytorch_int8
fix_pytorch_int8()


# import subprocess
# result = subprocess.run(['git', 'clone', 'https://huggingface.co/KumaTea/twitter-int8', 'model'], capture_output=True, text=True)
# print(result.stdout)


# Credit:
# https://huggingface.co/spaces/ljsabc/Fujisaki/blob/main/app.py


import torch
import logging
import gradio as gr
from transformers import AutoTokenizer, GenerationConfig, AutoModel


gr_title = """<h1 align="center">KumaGLM Lite</h1>
<h3 align='center'>这是<a href="https://huggingface.co/spaces/KumaTea/KumaGLM" target="_blank">另一个</a> AI Kuma，你可以与他聊天，或者直接在文本框按下Enter</h3>
<p align='center'>采用 INT4 量化，速度很慢，仅作备用</p>
<p align='center'>GitHub Repo: <a class="github-button" href="https://github.com/KumaTea/ChatGLM" aria-label="Star KumaTea/ChatGLM on GitHub">KumaTea/ChatGLM</a></p>
<script async defer src="https://buttons.github.io/buttons.js"></script>
"""
gr_footer =  """<p align='center'>
本项目基于
<a href='https://github.com/ljsabc/Fujisaki' target='_blank'>ljsabc/Fujisaki</a>
，模型采用
<a href='https://huggingface.co/THUDM/chatglm-6b' target='_blank'>THUDM/chatglm-6b</a>
。
</p>
<p align='center'>
<em>每天起床第一句！</em>
</p>"""
default_start = ["你是谁？", "我是 kuma"]


# device = torch.device('cpu')
# torch.cuda.current_device = lambda : device

logging.basicConfig(
    format='%(asctime)s %(levelname)-8s %(message)s',
    level=logging.INFO,
    datefmt='%m/%d %H:%M:%S')

model = AutoModel.from_pretrained(
    "KumaTea/twitter-int4",
    trust_remote_code=True,
    revision="e2aecb2"
).float()  # .to(device)
tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True, revision="4de8efe")

# dump a log to ensure everything works well
# print(model.peft_config)
# We have to use full precision, as some tokens are >65535
model.eval()
# print(model)

torch.set_default_tensor_type(torch.FloatTensor)


def evaluate(context, temperature, top_p, top_k=None):
    generation_config = GenerationConfig(
        temperature=temperature,
        top_p=top_p,
        # top_k=top_k,
        #repetition_penalty=1.1,
        num_beams=1,
        do_sample=True,
    )
    with torch.no_grad():
        # input_text = f"Context: {context}Answer: " 
        input_text = '||'.join(default_start) + '||'
        input_text += context + '||'
        logging.info('[API] Incoming request: ' + input_text)
        ids = tokenizer([input_text], return_tensors="pt")
        inputs = ids.to("cpu")
        out = model.generate(
            **inputs,
            max_length=224,
            generation_config=generation_config
        )
        out = out.tolist()[0]
        decoder_output = tokenizer.decode(out)
        # out_text = decoder_output.split("Answer: ")[1]
        out_text = decoder_output
        logging.info('[API] Result: ' + out_text)
        return out_text


def evaluate_stream(msg, history, temperature, top_p):
    generation_config = GenerationConfig(
        temperature=temperature,
        top_p=top_p,
        #repetition_penalty=1.1,
        num_beams=1,
        do_sample=True,
    )
    if not msg:
        msg = '……'

    history.append([msg, ""])

    context = '||'.join(default_start) + '||'
    if len(history) > 4:
        history.pop(0)

    for j in range(len(history)):
        history[j][0] = history[j][0].replace("<br>", "")

    # concatenate context
    for h in history[:-1]:
        context += h[0] + "||" + h[1] + "||"

    context += history[-1][0] + "||"
    context = context.replace(r'<br>', '')

    # TODO: Avoid the tokens are too long.
    CUTOFF = 224
    while len(tokenizer.encode(context)) > CUTOFF:
        # save 15 token size for the answer
        context = context[15:]

    h = []
    logging.info('[UI] Incoming request: ' + context)
    for response, h in model.stream_chat(tokenizer, context, h, max_length=CUTOFF, top_p=top_p, temperature=temperature):
        history[-1][1] = response
        yield history, ""


with gr.Blocks() as demo:
    gr.HTML(gr_title)
    # state = gr.State()
    with gr.Row():
        with gr.Column(scale=2):
            temp = gr.components.Slider(minimum=0, maximum=1.1, value=0.5, label="Temperature",
                info="温度参数，越高的温度生成的内容越丰富，但是有可能出现语法问题。小的温度也能帮助生成更相关的回答。")
            top_p = gr.components.Slider(minimum=0.5, maximum=1.0, value=0.8, label="Top-p",
                info="top-p参数，只输出前p>top-p的文字，越大生成的内容越丰富，但也可能出现语法问题。数字越小似乎上下文的衔接性越好。")
            #code = gr.Textbox(label="temp_output", info="解码器输出")
            #top_k = gr.components.Slider(minimum=1, maximum=200, step=1, value=25, label="Top k",
            #    info="top-k参数，下一个输出的文字会从top-k个文字中进行选择，越大生成的内容越丰富，但也可能出现语法问题。数字越小似乎上下文的衔接性越好。")
            
        with gr.Column(scale=3):
            chatbot = gr.Chatbot(label="聊天框", info="")
            msg = gr.Textbox(label="输入框", placeholder="最近过得怎么样？",
                info="输入你的内容，按 [Enter] 发送。什么都不填经常会出错。")
            clear = gr.Button("清除聊天")
            api_handler = gr.Button("API", visible=False)
            textbox_for_api = gr.Textbox(visible=False)

    msg.submit(evaluate_stream, [msg, chatbot, temp, top_p], [chatbot, msg])
    clear.click(lambda: None, None, chatbot, queue=False)
    api_handler.click(evaluate, [textbox_for_api, temp, top_p], [textbox_for_api], api_name='chat')
    gr.HTML(gr_footer)

demo.queue()
demo.launch(debug=False)