PyTorch 2.0 (original) (raw)

🤗 Diffusers supports the latest optimizations from PyTorch 2.0 which include:

1. A memory-efficient attention implementation, scaled dot product attention, without requiring any extra dependencies such as xFormers.
2. torch.compile, a just-in-time (JIT) compiler to provide an extra performance boost when individual models are compiled.

Both of these optimizations require PyTorch 2.0 or later and 🤗 Diffusers > 0.13.0.

pip install --upgrade torch diffusers

Scaled dot product attention

torch.nn.functional.scaled_dot_product_attention (SDPA) is an optimized and memory-efficient attention (similar to xFormers) that automatically enables several other optimizations depending on the model inputs and GPU type. SDPA is enabled by default if you’re using PyTorch 2.0 and the latest version of 🤗 Diffusers, so you don’t need to add anything to your code.

However, if you want to explicitly enable it, you can set a DiffusionPipeline to use AttnProcessor2_0:

import torch from diffusers import DiffusionPipeline

• from diffusers.models.attention_processor import AttnProcessor2_0

  pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).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]

SDPA should be as fast and memory efficient as xFormers; check the benchmark for more details.

In some cases - such as making the pipeline more deterministic or converting it to other formats - it may be helpful to use the vanilla attention processor, AttnProcessor. To revert to AttnProcessor, call the set_default_attn_processor() function on the pipeline:

import torch from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")

• pipe.unet.set_default_attn_processor()

  prompt = "a photo of an astronaut riding a horse on mars" image = pipe(prompt).images[0]

torch.compile

The torch.compile function can often provide an additional speed-up to your PyTorch code. In 🤗 Diffusers, it is usually best to wrap the UNet with torch.compile because it does most of the heavy lifting in the pipeline.

from diffusers import DiffusionPipeline import torch

pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda") 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[0]

Depending on GPU type, torch.compile can provide an additional speed-up of 5-300x on top of SDPA! If you’re using more recent GPU architectures such as Ampere (A100, 3090), Ada (4090), and Hopper (H100), torch.compile is able to squeeze even more performance out of these GPUs.

Compilation requires some time to complete, so it is best suited for situations where you prepare your pipeline once and then perform the same type of inference operations multiple times. For example, calling the compiled pipeline on a different image size triggers compilation again which can be expensive.

For more information and different options about torch.compile, refer to the torch_compile tutorial.

Learn more about other ways PyTorch 2.0 can help optimize your model in the Accelerate inference of text-to-image diffusion models tutorial.

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. The code is benchmarked on 🤗 Diffusers v0.17.0.dev0 to optimize torch.compile usage (see here for more details).

Expand the dropdown below to find the code used to benchmark each pipeline:

Stable Diffusion text-to-image

from diffusers import DiffusionPipeline import torch

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

run_compile = True

pipe = DiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True) 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 from diffusers.utils import load_image import torch

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

init_image = load_image(url) init_image = init_image.resize((512, 512))

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

run_compile = True

pipe = StableDiffusionImg2ImgPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True) 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 from diffusers.utils import load_image import torch

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 = load_image(img_url).resize((512, 512)) mask_image = load_image(mask_url).resize((512, 512))

path = "runwayml/stable-diffusion-inpainting"

run_compile = True

pipe = StableDiffusionInpaintPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True) 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 from diffusers.utils import load_image import torch

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

init_image = load_image(url) init_image = init_image.resize((512, 512))

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

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

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]

DeepFloyd IF text-to-image + upscaling

from diffusers import DiffusionPipeline import torch

run_compile = True

pipe_1 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True) pipe_1.to("cuda") pipe_2 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True) pipe_2.to("cuda") pipe_3 = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16, use_safetensors=True) pipe_3.to("cuda")

pipe_1.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_1.unet = torch.compile(pipe_1.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_1 = pipe_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images image_2 = pipe_2(image=image_1, prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images image_3 = pipe_3(prompt=prompt, image=image_1, noise_level=100).images

The graph below highlights the relative speed-ups for the StableDiffusionPipeline across five GPU families with PyTorch 2.0 and torch.compile enabled. The benchmarks for the following graphs are measured in number of iterations/second.

To give you an even better idea of how this speed-up holds for the other pipelines, consider the following graph for an A100 with PyTorch 2.0 and torch.compile:

In the following tables, we report our findings in terms of the number of iterations/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 DeepFloyd IF pipeline where 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. 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.

< > Update on GitHub