app.py

from diffusers import DiffusionPipeline, LCMScheduler, AutoencoderTiny
from compel import Compel, ReturnedEmbeddingsType
import torch
import os

from PIL import Image
import numpy as np
import gradio as gr
import psutil
from sfast.compilers.stable_diffusion_pipeline_compiler import (
    compile,
    CompilationConfig,
)


SAFETY_CHECKER = os.environ.get("SAFETY_CHECKER", None)
HF_TOKEN = os.environ.get("HF_TOKEN", None)
# check if MPS is available OSX only M1/M2/M3 chips
mps_available = hasattr(torch.backends, "mps") and torch.backends.mps.is_available()
xpu_available = hasattr(torch, "xpu") and torch.xpu.is_available()
device = torch.device(
    "cuda" if torch.cuda.is_available() else "xpu" if xpu_available else "cpu"
)
torch_device = device

#torch_dtype = torch.float16
torch_dtype = torch.bfloat16

print(f"SAFETY_CHECKER: {SAFETY_CHECKER}")
print(f"device: {device}")

if mps_available:
    device = torch.device("mps")
    torch_device = "cpu"
    torch_dtype = torch.float32

model_id = "stabilityai/stable-diffusion-xl-base-1.0"

if SAFETY_CHECKER == "True":
    pipe = DiffusionPipeline.from_pretrained(model_id)
else:
    pipe = DiffusionPipeline.from_pretrained(model_id, safety_checker=None)

pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe.load_lora_weights(
    "latent-consistency/lcm-lora-sdxl",
    use_auth_token=HF_TOKEN,
)
if device.type != "mps":
    pipe.unet.to(memory_format=torch.channels_last)
pipe.to(device=torch_device, dtype=torch_dtype).to(device)

# Load LCM LoRA

config = CompilationConfig.Default()
config.enable_xformers = False
config.enable_triton = False
config.enable_cuda_graph = False
pipe = compile(pipe, config=config)

compel_proc = Compel(
    tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
    text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
    requires_pooled=[False, True],
)


def predict(
    prompt,
    guidance,
    steps,
    seed=1231231,
    randomize_bt=False,
    progress=gr.Progress(track_tqdm=True),
):
    if randomize_bt:
        seed = np.random.randint(0, 2**32 - 1)
    generator = torch.manual_seed(seed)
    prompt_embeds, pooled_prompt_embeds = compel_proc(prompt)

    results = pipe(
        prompt_embeds=prompt_embeds,
        pooled_prompt_embeds=pooled_prompt_embeds,
        generator=generator,
        num_inference_steps=steps,
        guidance_scale=guidance,
        width=512,
        height=512,
        # original_inference_steps=params.lcm_steps,
        output_type="pil",
    )
    nsfw_content_detected = (
        results.nsfw_content_detected[0]
        if "nsfw_content_detected" in results
        else False
    )
    if nsfw_content_detected:
        raise gr.Error("NSFW content detected.")
    return results.images[0], seed


css = """
#container{
    margin: 0 auto;
    max-width: 40rem;
}
#intro{
    max-width: 100%;
    text-align: center;
    margin: 0 auto;
}
"""
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="container"):
        gr.Markdown(
            """# SDXL in 4 steps with Latent Consistency LoRAs
            SDXL is loaded with a LCM-LoRA, giving it the super power of doing inference in as little as 4 steps. [Learn more on our blog](https://huggingface.co/blog/lcm_lora) or [technical report](https://huggingface.co/papers/2311.05556).
            """,
            elem_id="intro",
        )
        with gr.Row():
            with gr.Row():
                prompt = gr.Textbox(
                    placeholder="Insert your prompt here:", scale=5, container=False
                )
                generate_bt = gr.Button("Generate", scale=1)

        image = gr.Image(type="filepath")
        with gr.Accordion("Advanced options", open=False):
            guidance = gr.Slider(
                label="Guidance", minimum=0.0, maximum=5, value=0.3, step=0.001
            )
            steps = gr.Slider(label="Steps", value=4, minimum=2, maximum=10, step=1)
            with gr.Row():
                seed = gr.Slider(
                    randomize=True,
                    minimum=0,
                    maximum=12013012031030,
                    label="Seed",
                    step=1,
                    scale=5,
                )
                with gr.Group():
                    randomize_bt = gr.Checkbox(label="Randomize", value=False)
                    random_seed = gr.Textbox(show_label=False)
        with gr.Accordion("Run with diffusers"):
            gr.Markdown(
                """## Running LCM-LoRAs it with `diffusers`
            ```bash
            pip install diffusers==0.23.0
            ```
            
            ```py
            from diffusers import DiffusionPipeline, LCMScheduler
            pipe = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0").to("cuda") 
            pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
            pipe.load_lora_weights("latent-consistency/lcm-lora-sdxl") #yes, it's a normal LoRA
            results = pipe(
                prompt="The spirit of a tamagotchi wandering in the city of Vienna",
                num_inference_steps=4,
                guidance_scale=0.0,
            )
            results.images[0]
            ```
            """
            )

        inputs = [prompt, guidance, steps, seed, randomize_bt]
        generate_bt.click(fn=predict, inputs=inputs, outputs=[image, random_seed])

demo.queue(api_open=False)
demo.launch(show_api=False)