app.py

# Reference:
# https://github.com/li-plus/chatglm.cpp
# https://github.com/li-plus/chatglm.cpp/blob/main/examples/web_demo.py

import chatglm_cpp
import gradio as gr
import argparse
from pathlib import Path

pipeline = chatglm_cpp.Pipeline("./chatglm3-ggml.bin")

max_length = 2048
top_p = 0.4
temp = 0.95
max_context_length=512
mode = "chat"
top_k = 0
repeat_penalty = 1.0
threads = 0

def postprocess(text):
    #if args.plain:
    #    return f"<pre>{text}</pre>"
    return text
    
def predict(input, chatbot, max_length, top_p, temperature, messages):
    chatbot.append((postprocess(input), ""))
    messages.append(chatglm_cpp.ChatMessage(role="user", content=input))

    generation_kwargs = dict(
        max_length=max_length,
        max_context_length=max_context_length,
        do_sample=temperature > 0,
        top_k=top_k,
        top_p=top_p,
        temperature=temperature,
        repetition_penalty=repeat_penalty,
        num_threads=threads,
        stream=True,
    )
    
    response = ""
    if mode == "chat":
        chunks = []
        for chunk in pipeline.chat(messages, **generation_kwargs):
            response += chunk.content
            chunks.append(chunk)
            chatbot[-1] = (chatbot[-1][0], postprocess(response))
            yield chatbot, messages
        messages.append(pipeline.merge_streaming_messages(chunks))
    else:
        for chunk in pipeline.generate(input, **generation_kwargs):
            response += chunk
            chatbot[-1] = (chatbot[-1][0], postprocess(response))
            yield chatbot, messages

    yield chatbot, messages

def reset_user_input():
    return gr.update(value="")


def reset_state():
    return [], []

with gr.Blocks() as demo:
    gr.HTML("""<h1 align="center">ChatGLM3 Quantized by ChatGLM.cpp. Reduce size from 12G to 3.4G.</h1>""")

    chatbot = gr.Chatbot()
    with gr.Row():
        with gr.Column(scale=4):
            user_input = gr.Textbox(show_label=False, placeholder="Input...", lines=8)
            submitBtn = gr.Button("Submit", variant="primary")
        with gr.Column(scale=1):
            max_length = gr.Slider(0, 2048, value=max_length, step=1.0, label="Maximum Length", interactive=True)
            top_p = gr.Slider(0, 1, value=top_p, step=0.01, label="Top P", interactive=True)
            temperature = gr.Slider(0, 1, value=temp, step=0.01, label="Temperature", interactive=True)
            emptyBtn = gr.Button("Clear History")

    messages = gr.State([])

    submitBtn.click(
        predict,
        [user_input, chatbot, max_length, top_p, temperature, messages],
        [chatbot, messages],
        show_progress=True,
    )
    submitBtn.click(reset_user_input, [], [user_input])

    emptyBtn.click(reset_state, outputs=[chatbot, messages], show_progress=True)

demo.queue().launch(share=False, inbrowser=True)