app.py

import psutil
import gradio as gr

from functools import partial
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

mem = psutil.virtual_memory()

tokenizer = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)

model = AutoModelForSeq2SeqLM.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).bfloat16()

def chat(query, history=[]):
    _, history = model.chat(tokenizer, query, history, max_length=512)
    return history, history

description = "ChatGLM-6B 是一个开源的、支持中英双语问答的对话语言模型，基于 General Language Model (GLM) 架构，具有 62 亿参数。 \n"
title = "ChatGLM-6B Chatbot"
examples = [["Hello"], ["你好"], ["介绍清华"]]

chatbot_interface = gr.Interface(
    fn=chat,
    title=title,
    description=description,
    examples=examples,
    inputs=["text", "state"],
    outputs=["chatbot", "state"]
)

# chatbot_interface.launch()