app.py

import gradio as gr
import requests

def generate(prompt = "", model = "bigscience/bloom", use_cache = False, **kwargs):
  json = {"inputs": prompt, "parameters": kwargs, "options": {"use_cache": use_cache}}
  response = requests.post(f"https://api-inference.huggingface.co/models/{model}", json=json)
  try:
    output = response.json()[0]["generated_text"].replace(prompt, "")
  except:
    output = ""
  return output

def generate_line(story, temperature):
  out = generate(story, temperature=temperature, max_new_tokens=50)
  i = out.find("\n")
  return story + out[:i]

with gr.Blocks() as demo:
  story = gr.Textbox(value="Character_1:\n\nCharacter_2:", label="Story")
  generate_button = gr.Button("Generate")
  generate_button.click(generate_line, inputs=[story, gr.Slider(0, 2, 0.9, label="temperature")], outputs=[story])
  
demo.launch()