app.py

import os
import random
import uuid
import json

import gradio as gr
import numpy as np
from PIL import Image
import spaces
import torch
from diffusers import StableDiffusionXLPipeline, EulerAncestralDiscreteScheduler

# Use environment variables for flexibility
MODEL_ID = os.getenv("MODEL_ID", "sd-community/sdxl-flash")
MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "4096"))
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE", "0") == "1"
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD", "0") == "1"
BATCH_SIZE = int(os.getenv("BATCH_SIZE", "1"))  # Allow generating multiple images at once

# Determine device and load model outside of function for efficiency
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
pipe = StableDiffusionXLPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
    use_safetensors=True,
    add_watermarker=False,
).to(device)
pipe.scheduler = EulerAncestralDiscreteScheduler.from_config(pipe.scheduler.config)

# Torch compile for potential speedup (experimental)
if USE_TORCH_COMPILE:
    pipe.compile()

# CPU offloading for larger RAM capacity (experimental)
if ENABLE_CPU_OFFLOAD:
    pipe.enable_model_cpu_offload()

MAX_SEED = np.iinfo(np.int32).max

def save_image(img):
    unique_name = str(uuid.uuid4()) + ".png"
    img.save(unique_name)
    return unique_name

def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    return seed

@spaces.GPU(duration=35, enable_queue=True)
def generate(
    prompt: str,
    negative_prompt: str = "",
    use_negative_prompt: bool = False,
    seed: int = 1,
    width: int = 1024,
    height: int = 1024,
    guidance_scale: float = 3,
    num_inference_steps: int = 30,
    randomize_seed: bool = False,
    use_resolution_binning: bool = True, 
    num_images: int = 1,  # Number of images to generate
    progress=gr.Progress(track_tqdm=True),
):
    seed = int(randomize_seed_fn(seed, randomize_seed))
    generator = torch.Generator(device=device).manual_seed(seed)

    # Improved options handling
    options = {
        "prompt": [prompt] * num_images,
        "negative_prompt": [negative_prompt] * num_images if use_negative_prompt else None,
        "width": width,
        "height": height,
        "guidance_scale": guidance_scale,
        "num_inference_steps": num_inference_steps,
        "generator": generator,
        "output_type": "pil",
    }

    # Use resolution binning for faster generation with less VRAM usage
    if use_resolution_binning:
        options["use_resolution_binning"] = True

    # Generate images potentially in batches
    images = []
    for i in range(0, num_images, BATCH_SIZE):
        batch_options = options.copy()
        batch_options["prompt"] = options["prompt"][i:i+BATCH_SIZE]
        if "negative_prompt" in batch_options:
            batch_options["negative_prompt"] = options["negative_prompt"][i:i+BATCH_SIZE]
        images.extend(pipe(**batch_options).images)

    image_paths = [save_image(img) for img in images]
    return image_paths, seed

examples = [
    "a cat eating a piece of cheese",
    "a ROBOT riding a BLUE horse on Mars, photorealistic, 4k",
    "Ironman VS Hulk, ultrarealistic",
    "Astronaut in a jungle, cold color palette, oil pastel, detailed, 8k",
    "An alien holding a sign board containing the word 'Flash', futuristic, neonpunk",
    "Kids going to school, Anime style"
]

css = '''
.gradio-container{max-width: 700px !important}
h1{text-align:center}
footer {
    visibility: hidden
}

.wheel-and-hamster {
  --dur: 1s;
  position: relative;
  width: 12em;
  height: 12em;
  font-size: 14px;
}

.wheel,
.hamster,
.hamster div,
.spoke {
  position: absolute;
}

.wheel,
.spoke {
  border-radius: 50%;
  top: 0;
  left: 0;
  width: 100%;
  height: 100%;
}

.wheel {
  background: radial-gradient(100% 100% at center,hsla(0,0%,60%,0) 47.8%,hsl(0,0%,60%) 48%);
  z-index: 2;
}

.hamster {
  animation: hamster var(--dur) ease-in-out infinite;
  top: 50%;
  left: calc(50% - 3.5em);
  width: 7em;
  height: 3.75em;
  transform: rotate(4deg) translate(-0.8em,1.85em);
  transform-origin: 50% 0;
  z-index: 1;
}

.hamster__head {
  animation: hamsterHead var(--dur) ease-in-out infinite;
  background: hsl(30,90%,55%);
  border-radius: 70% 30% 0 100% / 40% 25% 25% 60%;
  box-shadow: 0 -0.25em 0 hsl(30,90%,80%) inset,
		0.75em -1.55em 0 hsl(30,90%,90%) inset;
  top: 0;
  left: -2em;
  width: 2.75em;
  height: 2.5em;
  transform-origin: 100% 50%;
}

.hamster__ear {
  animation: hamsterEar var(--dur) ease-in-out infinite;
  background: hsl(0,90%,85%);
  border-radius: 50%;
  box-shadow: -0.25em 0 hsl(30,90%,55%) inset;
  top: -0.25em;
  right: -0.25em;
  width: 0.75em;
  height: 0.75em;
  transform-origin: 50% 75%;
}

.hamster__eye {
  animation: hamsterEye var(--dur) linear infinite;
  background-color: hsl(0,0%,0%);
  border-radius: 50%;
  top: 0.375em;
  left: 1.25em;
  width: 0.5em;
  height: 0.5em;
}

.hamster__nose {
  background: hsl(0,90%,75%);
  border-radius: 35% 65% 85% 15% / 70% 50% 50% 30%;
  top: 0.75em;
  left: 0;
  width: 0.2em;
  height: 0.25em;
}

.hamster__body {
  animation: hamsterBody var(--dur) ease-in-out infinite;
  background: hsl(30,90%,90%);
  border-radius: 50% 30% 50% 30% / 15% 60% 40% 40%;
  box-shadow: 0.1em 0.75em 0 hsl(30,90%,55%) inset,
		0.15em -0.5em 0 hsl(30,90%,80%) inset;
  top: 0.25em;
  left: 2em;
  width: 4.5em;
  height: 3em;
  transform-origin: 17% 50%;
  transform-style: preserve-3d;
}

.hamster__limb--fr,
.hamster__limb--fl {
  clip-path: polygon(0 0,100% 0,70% 80%,60% 100%,0% 100%,40% 80%);
  top: 2em;
  left: 0.5em;
  width: 1em;
  height: 1.5em;
  transform-origin: 50% 0;
}

.hamster__limb--fr {
  animation: hamsterFRLimb var(--dur) linear infinite;
  background: linear-gradient(hsl(30,90%,80%) 80%,hsl(0,90%,75%) 80%);
  transform: rotate(15deg) translateZ(-1px);
}

.hamster__limb--fl {
  animation: hamsterFLLimb var(--dur) linear infinite;
  background: linear-gradient(hsl(30,90%,80%) 80%,hsl(0,90%,75%) 80%);
  transform: rotate(-60deg) translateZ(-1px);
}

.hamster__limb--br,
.hamster__limb--bl {
  clip-path: polygon(0 0,100% 0,100% 20%,30% 100%,0% 100%);
  top: 1.25em;
  left: 2.8em;
  width: 1.5em;
  height: 2.5em;
  transform-origin: 33% 10%;
}

.hamster__limb--br {
  animation: hamsterBRLimb var(--dur) linear infinite;
  background: linear-gradient(hsl(0,90%,75%) 40%,hsl(30,90%,80%) 40%);
  transform: rotate(-15deg) translateZ(-1px);
}

.hamster__limb--bl {
  animation: hamsterBLLimb var(--dur) linear infinite;
  background: linear-gradient(hsl(0,90%,75%) 40%,hsl(30,90%,80%) 40%);
  transform: rotate(60deg) translateZ(-1px);
}

.hamster__tail {
  animation: hamsterTail var(--dur) linear infinite;
  background: hsl(0,90%,85%);
  border-radius: 0.25em 50% 50% 0.25em;
  box-shadow: 0.25em 0 hsl(30,90%,55%) inset;
  top: 1.5em;
  left: 5.5em;
  width: 0.5em;
  height: 0.75em;
  transform: rotate(30deg) translateZ(-1px);
  transform-origin: 0.25em 0.125em;
}

.spoke {
  background: radial-gradient(hsl(0,0%,70%) 25%,hsla(0,0%,60%,0) 26%) center/8px 8px;
  z-index: 0;
}

.spoke--1 {
  animation: spoke var(--dur) linear infinite;
}

.spoke--2 {
  animation: spoke var(--dur) linear infinite;
  transform: rotate(30deg);
}

.spoke--3 {
  animation: spoke var(--dur) linear infinite;
  transform: rotate(60deg);
}

@keyframes hamster {
  0%,100% { transform: rotate(4deg) translate(-0.8em,1.85em) }
  50% { transform: rotate(0) translate(-0.8em,1.85em) }
}

@keyframes hamsterHead {
  0%,100% { transform: rotate(0) }
  50% { transform: rotate(-8deg) }
}

@keyframes hamsterEar {
  0%,100% { transform: rotate(0) }
  50% { transform: rotate(-3deg) }
}

@keyframes hamsterEye {
  0%,90%,100% { transform: scaleY(1) }
  95% { transform: scaleY(0) }
}

@keyframes hamsterBody {
  0%,100% { transform: rotate(0) }
  50% { transform: rotate(2deg) }
}

@keyframes hamsterFRLimb {
  0%,100% { transform: rotate(15deg) translateZ(-1px) }
  50% { transform: rotate(-30deg) translateZ(-1px) }
}

@keyframes hamsterFLLimb {
  0%,100% { transform: rotate(-60deg) translateZ(-1px) }
  50% { transform: rotate(-25deg) translateZ(-1px) }
}

@keyframes hamsterBRLimb {
  0%,100% { transform: rotate(-15deg) translateZ(-1px) }
  50% { transform: rotate(30deg) translateZ(-1px) }
}

@keyframes hamsterBLLimb {
  0%,100% { transform: rotate(60deg) translateZ(-1px) }
  50% { transform: rotate(25deg) translateZ(-1px) }
}

@keyframes hamsterTail {
  0%,100% { transform: rotate(30deg) translateZ(-1px) }
  50% { transform: rotate(10deg) translateZ(-1px) }
}

@keyframes spoke {
  0% { transform: rotate(0) }
  100% { transform: rotate(1turn) }
}
'''

html = '''
<div id="loading-animation" style="display: flex; justify-content: center; align-items: center; height: 100vh;">
  <div class="wheel-and-hamster">
    <div class="wheel"></div>
    <div class="hamster">
      <div class="hamster__body">
        <div class="hamster__head">
          <div class="hamster__ear"></div>
          <div class="hamster__eye"></div>
          <div class="hamster__nose"></div>
        </div>
        <div class="hamster__limb hamster__limb--fr"></div>
        <div class="hamster__limb hamster__limb--fl"></div>
        <div class="hamster__limb hamster__limb--br"></div>
        <div class="hamster__limb hamster__limb--bl"></div>
        <div class="hamster__tail"></div>
      </div>
    </div>
    <div class="spoke spoke--1"></div>
    <div class="spoke spoke--2"></div>
    <div class="spoke spoke--3"></div>
  </div>
</div>
<script>
  window.onload = function() {
    document.getElementById("loading-animation").style.display = "none";
  }
</script>
'''

with gr.Blocks(css=css) as demo:
    gr.HTML(html)
    gr.Markdown("# Flash Attention with SDXL")
    gr.Markdown("Generate images with Flash Attention and SDXL")

    with gr.Row():
        with gr.Column(scale=55):
            prompt = gr.Textbox(label="Prompt", show_label=False, max_lines=2, placeholder="Enter your prompt").style(container=False)
            negative_prompt = gr.Textbox(label="Negative Prompt", show_label=False, max_lines=2, placeholder="Enter negative prompt").style(container=False)
        with gr.Column(scale=45):
            generate_btn = gr.Button("Generate")

    with gr.Row():
        image_output = gr.Gallery(label="Generated Images").style(grid=2, height="auto")
        seed_output = gr.Number(label="Seed Used")

    gr.Examples(examples=examples, inputs=[prompt])

    inputs = [prompt, negative_prompt, gr.Checkbox(False, label="Use Negative Prompt"), gr.Slider(1, MAX_SEED, value=1, label="Seed"),
              gr.Slider(256, MAX_IMAGE_SIZE, value=1024, label="Width"), gr.Slider(256, MAX_IMAGE_SIZE, value=1024, label="Height"),
              gr.Slider(1, 20, value=7.5, label="Guidance Scale"), gr.Slider(1, 100, value=30, label="Number of Inference Steps"),
              gr.Checkbox(False, label="Randomize Seed"), gr.Checkbox(True, label="Use Resolution Binning"),
              gr.Slider(1, 10, value=1, label="Number of Images")]

    generate_btn.click(fn=generate, inputs=inputs, outputs=[image_output, seed_output])

demo.launch()