app.py

import gradio as gr

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer, pipeline
from threading import Thread

# The huggingface model id for Microsoft's phi-2 model
checkpoint = "microsoft/phi-2"

# Download and load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.float32, device_map="cpu", trust_remote_code=True)

# Text generation pipeline
phi2 = pipeline(
    "text-generation", 
    tokenizer=tokenizer, 
    model=model,  
    pad_token_id=tokenizer.eos_token_id,
    eos_token_id=tokenizer.eos_token_id,
    device_map="cpu"
)

# Function that accepts a prompt and generates text using the phi2 pipeline
def generate(message, chat_history, max_new_tokens):

  instruction = "You are a helpful assistant to 'User'. You do not respond as 'User' or pretend to be 'User'. You only respond once as 'Assistant'."
  final_prompt = f"Instruction: {instruction}\n"

  for sent, received in chat_history:
    final_prompt += "User: " + sent + "\n"
    final_prompt += "Assistant: " + received + "\n"

  final_prompt += "User: " + message + "\n"
  final_prompt += "Output:"

  if len(tokenizer.tokenize(final_prompt)) >= tokenizer.model_max_length - max_new_tokens:
    final_prompt = "Instruction: Say 'Input exceeded context size, please clear the chat history and retry!' Output:"
    
  # Streamer
  streamer = TextIteratorStreamer(tokenizer=tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=300.0)
  thread = Thread(target=phi2, kwargs={"text_inputs":final_prompt, "max_new_tokens":max_new_tokens, "streamer":streamer})
  thread.start()

  generated_text = ""
  for word in streamer:
    generated_text += word
    response = generated_text.strip()

    if "User:" in response:
      response = response.split("User:")[0].strip()

    if "Assistant:" in response:
      response = response.split("Assistant:")[1].strip()

    yield response

# Chat interface with gradio
with gr.Blocks() as demo:
  gr.Markdown("""
  # Phi-2 Chatbot Demo
  This chatbot was created using Microsoft's 2.7 billion parameter [phi-2](https://huggingface.co/microsoft/phi-2) Transformer model. 
  
  In order to reduce the response time on this hardware, `max_new_tokens` has been set to `21` in the text generation pipeline. With this default configuration, it takes approximately `60 seconds` for the response to start being generated, and streamed one word at a time. Use the slider below to increase or decrease the length of the generated text.
  """)

  tokens_slider = gr.Slider(8, 128, value=21, label="Maximum new tokens", info="A larger `max_new_tokens` parameter value gives you longer text responses but at the cost of a slower response time.")

  chatbot = gr.ChatInterface(
    fn=generate,
    additional_inputs=[tokens_slider],
    stop_btn=None,
    examples=[["Who is Leonhard Euler?"]]
  )
  
demo.queue().launch()