app.py

import gradio as gr
from gradio.inputs import Textbox, Slider

import requests

# Template
title = "A conversation with some NPC in a Tavern 🍻"
description = ""
article = """
<p> If you liked don't forget to 💖 the project 🥰 </p>
<h2> Parameters: </h2>
<ul>
    <li><i>message</i>: what you want to say to the NPC.</li>
    <li><i>npc_name</i>: name of the NPC.</li>
    <li><i>npc_prompt</i>: prompt of the NPC, we can modify it to see if results are better.</li>
    <li><i>top_p</i>:  control how deterministic the model is in generating a response.</li>
    <li><i>temperature</i>: (sampling temperature) higher values means the model will take more risks.</li>
    <li><i>max_new_tokens</i>: Max number of tokens in generation.</li>
</ul>
<img src='http://www.simoninithomas.com/test/gandalf.jpg', alt="Gandalf"/>"""
theme="huggingface"


# Builds the prompt from what previously happened 
def build_prompt(conversation, context, interlocutor_names):
  prompt = context + "\n"
  for player_msg, npc_msg in conversation:
      line = "\n- " + interlocutor_names[0] + ":" + player_msg
      prompt += line
      line = "\n- " + interlocutor_names[1] + ":" + npc_msg
      prompt += line
  prompt += ""
  return prompt

# Recognize what the model said, if it used the correct format
def clean_chat_output(txt, prompt, interlocutor_names):
  delimiter = "\n- "+interlocutor_names[0]
  output = txt.replace(prompt, '')
  output = output[:output.find(delimiter)]
  return output

# GPT-J-6B API
API_URL = "https://api-inference.huggingface.co/models/EleutherAI/gpt-j-6B"
def query(payload):
  response = requests.post(API_URL, json=payload)
  return response.json()

def chat(message, npc_name, initial_prompt, top_p, temperature, max_new_tokens, history=[]):
    interlocutor_names = ["Player", npc_name]
    
    print("message", message)
    print("npc_name", npc_name)
    print("initial_prompt", initial_prompt)
    print("top_p", top_p)
    print("temperature", temperature)
    print("max_new_tokens", max_new_tokens)
    print("history", history)
    response = "Test"
    history.append((message, ""))
    conversation = history
    
    # Build the prompt
    prompt = build_prompt(conversation, initial_prompt, interlocutor_names)
    
    # Build JSON
    json_req = {"inputs": prompt,
         "parameters":
         {
         "top_p": top_p,
        "temperature": temperature,
        "max_new_tokens": max_new_tokens,
        "return_full_text": False
        }}
    
    # Get the output
    output = query(json_req)
    output = output[0]['generated_text']
    print("output", output)
    
    answer = clean_chat_output(output, prompt, interlocutor_names)
    response = answer
    print("response", answer)
    history[-1] = (message, response)
    return history, history


#io = gr.Interface.load("huggingface/EleutherAI/gpt-j-6B")

iface = gr.Interface(fn=chat, 
inputs=[Textbox(label="message", placeholder="Hello!"),
        Textbox(label="npc_name", placeholder="Antoine"),
        Textbox(label="initial_prompt", placeholder="The following is a conversation with Antoine, a guard for Northfall that's drinking in the Tavern."),
        Slider(minimum=0.5, maximum=1, step=0.05, default=0.9, label="top_p"),
        Slider(minimum=0.5, maximum=1.5, step=0.1, default=1.1, label="temperature"),
        Slider(minimum=20, maximum=250, step=10, default=50, label="max_new_tokens"),
        "state"],
        outputs=["chatbot","state"],
        #examples = [["Hello!", "", , 0.9, 1.1, 50, iface.state]],
        allow_screenshot=True, 
        allow_flagging=True,
        title=title,
        article=article,
        theme=theme)
 
if __name__ == "__main__":
  iface.launch()