Accelerated PyTorch 2.0 support in Diffusers

Starting from version 0.13.0, Diffusers supports the latest optimization from PyTorch 2.0. These include:

1. Support for accelerated transformers implementation with memory-efficient attention – no extra dependencies (such as xformers) required.
2. torch.compile support for extra performance boost when individual models are compiled.

Installation

To benefit from the accelerated attention implementation and torch.compile(), you just need to install the latest versions of PyTorch 2.0 from pip, and make sure you are on diffusers 0.13.0 or later. As explained below, diffusers automatically uses the optimized attention processor (AttnProcessor2_0) (but not torch.compile()) when PyTorch 2.0 is available.

pip install --upgrade torch diffusers

Using accelerated transformers and torch.compile.

1. Accelerated Transformers implementation

   PyTorch 2.0 includes an optimized and memory-efficient attention implementation through the torch.nn.functional.scaled_dot_product_attention function, which automatically enables several optimizations depending on the inputs and the GPU type. This is similar to the memory_efficient_attention from xFormers, but built natively into PyTorch.

   These optimizations will be enabled by default in Diffusers if PyTorch 2.0 is installed and if torch.nn.functional.scaled_dot_product_attention is available. To use it, just install torch 2.0 as suggested above and simply use the pipeline. For example:

   Copied

   import torch
   from diffusers import DiffusionPipeline
   
   pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
   pipe = pipe.to("cuda")
   
   prompt = "a photo of an astronaut riding a horse on mars"
   image = pipe(prompt).images[0]

   If you want to enable it explicitly (which is not required), you can do so as shown below.

   Copied

   import torch
   from diffusers import DiffusionPipeline
   + from diffusers.models.attention_processor import AttnProcessor2_0
   
   pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
   + pipe.unet.set_attn_processor(AttnProcessor2_0())
   
   prompt = "a photo of an astronaut riding a horse on mars"
   image = pipe(prompt).images[0]

   This should be as fast and memory efficient as xFormers. More details in our benchmark.

   It is possible to revert to the vanilla attention processor (AttnProcessor), which can be helpful to make the pipeline more deterministic, or if you need to convert a fine-tuned model to other formats such as Core ML. To use the normal attention processor you can use the set_default_attn_processor() function:

   Copied

   import torch
   from diffusers import DiffusionPipeline
   from diffusers.models.attention_processor import AttnProcessor
   
   pipe = DiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
   pipe.unet.set_default_attn_processor()
   
   prompt = "a photo of an astronaut riding a horse on mars"
   image = pipe(prompt).images[0]

2. torch.compile

   To get an additional speedup, we can use the new torch.compile feature. Since the UNet of the pipeline is usually the most computationally expensive, we wrap the unet with torch.compile leaving rest of the sub-models (text encoder and VAE) as is. For more information and different options, refer to the torch compile docs.

   Copied

   pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
   images = pipe(prompt, num_inference_steps=steps, num_images_per_prompt=batch_size).images

   Depending on the type of GPU, compile() can yield between 5% - 300% of additional speed-up over the accelerated transformer optimizations. Note, however, that compilation is able to squeeze more performance improvements in more recent GPU architectures such as Ampere (A100, 3090), Ada (4090) and Hopper (H100).

   Compilation takes some time to complete, so it is best suited for situations where you need to prepare your pipeline once and then perform the same type of inference operations multiple times. Calling the compiled pipeline on a different image size will re-trigger compilation which can be expensive.

Benchmark

We conducted a comprehensive benchmark with PyTorch 2.0’s efficient attention implementation and torch.compile across different GPUs and batch sizes for five of our most used pipelines. We used diffusers 0.17.0.dev0, which makes sure torch.compile() is leveraged optimally.

Benchmarking code

Stable Diffusion text-to-image

from diffusers import DiffusionPipeline
import torch

path = "runwayml/stable-diffusion-v1-5"

run_compile = True  # Set True / False

pipe = DiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
pipe.unet.to(memory_format=torch.channels_last)

if run_compile:
    print("Run torch compile")
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

prompt = "ghibli style, a fantasy landscape with castles"

for _ in range(3):
    images = pipe(prompt=prompt).images

Stable Diffusion image-to-image

from diffusers import StableDiffusionImg2ImgPipeline
import requests
import torch
from PIL import Image
from io import BytesIO

url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"

response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))

path = "runwayml/stable-diffusion-v1-5"

run_compile = True  # Set True / False

pipe = StableDiffusionImg2ImgPipeline.from_pretrained(path, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
pipe.unet.to(memory_format=torch.channels_last)

if run_compile:
    print("Run torch compile")
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

prompt = "ghibli style, a fantasy landscape with castles"

for _ in range(3):
    image = pipe(prompt=prompt, image=init_image).images[0]

Stable Diffusion - inpainting

from diffusers import StableDiffusionInpaintPipeline
import requests
import torch
from PIL import Image
from io import BytesIO

url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"

def download_image(url):
    response = requests.get(url)
    return Image.open(BytesIO(response.content)).convert("RGB")


img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"

init_image = download_image(img_url).resize((512, 512))
mask_image = download_image(mask_url).resize((512, 512))

path = "runwayml/stable-diffusion-inpainting"

run_compile = True  # Set True / False

pipe = StableDiffusionInpaintPipeline.from_pretrained(path, torch_dtype=torch.float16)
pipe = pipe.to("cuda")
pipe.unet.to(memory_format=torch.channels_last)

if run_compile:
    print("Run torch compile")
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)

prompt = "ghibli style, a fantasy landscape with castles"

for _ in range(3):
    image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]

ControlNet

from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
import requests
import torch
from PIL import Image
from io import BytesIO

url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"

response = requests.get(url)
init_image = Image.open(BytesIO(response.content)).convert("RGB")
init_image = init_image.resize((512, 512))

path = "runwayml/stable-diffusion-v1-5"

run_compile = True  # Set True / False
controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16)
pipe = StableDiffusionControlNetPipeline.from_pretrained(
    path, controlnet=controlnet, torch_dtype=torch.float16
)

pipe = pipe.to("cuda")
pipe.unet.to(memory_format=torch.channels_last)
pipe.controlnet.to(memory_format=torch.channels_last)

if run_compile:
    print("Run torch compile")
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
    pipe.controlnet = torch.compile(pipe.controlnet, mode="reduce-overhead", fullgraph=True)

prompt = "ghibli style, a fantasy landscape with castles"

for _ in range(3):
    image = pipe(prompt=prompt, image=init_image).images[0]

IF text-to-image + upscaling

from diffusers import DiffusionPipeline
import torch

run_compile = True  # Set True / False

pipe = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16)
pipe.to("cuda")
pipe_2 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16)
pipe_2.to("cuda")
pipe_3 = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16)
pipe_3.to("cuda")


pipe.unet.to(memory_format=torch.channels_last)
pipe_2.unet.to(memory_format=torch.channels_last)
pipe_3.unet.to(memory_format=torch.channels_last)

if run_compile:
    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
    pipe_2.unet = torch.compile(pipe_2.unet, mode="reduce-overhead", fullgraph=True)
    pipe_3.unet = torch.compile(pipe_3.unet, mode="reduce-overhead", fullgraph=True)

prompt = "the blue hulk"

prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)
neg_prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)

for _ in range(3):
    image = pipe(prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
    image_2 = pipe_2(image=image, prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
    image_3 = pipe_3(prompt=prompt, image=image, noise_level=100).images

To give you a pictorial overview of the possible speed-ups that can be obtained with PyTorch 2.0 and torch.compile(), here is a plot that shows relative speed-ups for the Stable Diffusion text-to-image pipeline across five different GPU families (with a batch size of 4):

To give you an even better idea of how this speed-up holds for the other pipelines presented above, consider the following plot that shows the benchmarking numbers from an A100 across three different batch sizes (with PyTorch 2.0 nightly and torch.compile()):

(Our benchmarking metric for the plots above is number of iterations/second)

But we reveal all the benchmarking numbers in the interest of transparency!

In the following tables, we report our findings in terms of the number of iterations processed per second.

A100 (batch size: 1)

A100 (batch size: 4)

A100 (batch size: 16)

V100 (batch size: 1)

V100 (batch size: 4)

V100 (batch size: 16)

T4 (batch size: 1)

T4 (batch size: 4)

T4 (batch size: 16)

RTX 3090 (batch size: 1)

RTX 3090 (batch size: 4)

RTX 3090 (batch size: 16)

RTX 4090 (batch size: 1)

RTX 4090 (batch size: 4)

RTX 4090 (batch size: 16)

Notes

• Follow this PR for more details on the environment used for conducting the benchmarks.
• For the IF pipeline and batch sizes > 1, we only used a batch size of >1 in the first IF pipeline for text-to-image generation and NOT for upscaling. So, that means the two upscaling pipelines received a batch size of 1.

Thanks to Horace He from the PyTorch team for their support in improving our support of torch.compile() in Diffusers.