app.py

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import gradio as gr

tokenizer = AutoTokenizer.from_pretrained("microsoft/DialoGPT-large")
model = AutoModelForCausalLM.from_pretrained("microsoft/DialoGPT-large")

def predict(input, history=[]):
  new_user_input_ids = tokenizer.encode(input + tokenizer.eos_token, return_tensors='pt')
  bot_input_ids = torch.cat([torch.LongTensor(history), new_user_input_ids], dim=-1)
  history = model.generate(bot_input_ids, max_length=500, pad_token_id=tokenizer.eos_token_id).tolist()
  response = tokenizer.decode(history[0]).replace("<|endoftext|>", "\n")
  
  html = "<div class='chatbot'>"
  for m, msg in enumerate(response):
    cls = "user" if m%2 == 0 else "bot"
    html += "<div class='msg {}'> {}</div>".format(cls, msg)
  html += "</div>"
  
  return response, history

css = """
.chatbox {display:flex;flex-direction:column}
.msg {padding:4px;margin-bottom:4px;border-radius:4px;width:80%}
.msg.user {background-color:cornflowerblue;color:white}
.msg.bot {background-color:green;align-self:self-end}
.footer {display:none !important}
"""

gr.Interface(fn=predict, theme="grass", css=css, title="Chatbot Two", 
inputs=[gr.inputs.Textbox(placeholder="Write a text message as if writing a text message to a human."), "state"], outputs=["html", "state"]).launch()