app.py

import os
import torch
import gradio as gr
from torch import autocast
from diffusers import StableDiffusionPipeline

# get hf user access token as an environment variable
TOKEN_KEY = os.getenv('AUTH_TOKEN')

# choose GPU else fallback to CPU
device = "cuda" if torch.cuda.is_available() else "cpu"
device_name = torch.cuda.get_device_name(0)

# setup pipeline
pipe = StableDiffusionPipeline.from_pretrained("CompVis/stable-diffusion-v1-4", revision="fp16", torch_dtype=torch.float16, use_auth_token=TOKEN_KEY)
pipe = pipe.to(device)

# define gradio function
def generate(prompt:str, seed:int, guidance:float):
    generator = torch.Generator(device).manual_seed(int(seed))
    with autocast(device):
        image = pipe(prompt=prompt, generator=generator, guidance_scale=guidance, steps=50).images[0]
    return image

if device == "cuda":
    print(device_name + " available.")
    # create the gradio UI
    demo = gr.Interface(
        fn=generate,
        inputs=[gr.Textbox(placeholder="castle on a mountain"), gr.Number(value=123456), gr.Slider(0,10)],
        outputs="image",
        allow_flagging="never",
    )

    # allow queueing or incoming requests, max=3 
    demo.queue(concurrency_count=3)

    # launch demo
    demo.launch()
else:
    print("GPU unavailable.")