Hugging Face's logo

Diffusers documentation

Parallel Sampling of Diffusion Models (ParaDiGMS)

You are viewing v0.18.2 version. A newer version v0.27.2 is available.
Hugging Face's logo
Join the Hugging Face community

and get access to the augmented documentation experience

to get started

Parallel Sampling of Diffusion Models (ParaDiGMS)

Overview

Parallel Sampling of Diffusion Models by Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, Nima Anari.

The abstract of the paper is the following:

Diffusion models are powerful generative models but suffer from slow sampling, often taking 1000 sequential denoising steps for one sample. As a result, considerable efforts have been directed toward reducing the number of denoising steps, but these methods hurt sample quality. Instead of reducing the number of denoising steps (trading quality for speed), in this paper we explore an orthogonal approach: can we run the denoising steps in parallel (trading compute for speed)? In spite of the sequential nature of the denoising steps, we show that surprisingly it is possible to parallelize sampling via Picard iterations, by guessing the solution of future denoising steps and iteratively refining until convergence. With this insight, we present ParaDiGMS, a novel method to accelerate the sampling of pretrained diffusion models by denoising multiple steps in parallel. ParaDiGMS is the first diffusion sampling method that enables trading compute for speed and is even compatible with existing fast sampling techniques such as DDIM and DPMSolver. Using ParaDiGMS, we improve sampling speed by 2-4x across a range of robotics and image generation models, giving state-of-the-art sampling speeds of 0.2s on 100-step DiffusionPolicy and 16s on 1000-step StableDiffusion-v2 with no measurable degradation of task reward, FID score, or CLIP score.

Resources:

Available Pipelines:

Pipeline Tasks Demo
StableDiffusionParadigmsPipeline Faster Text-to-Image Generation

This pipeline was contributed by AndyShih12 in this PR.

Usage example 

import torch
from diffusers import DDPMParallelScheduler
from diffusers import StableDiffusionParadigmsPipeline

scheduler = DDPMParallelScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler")

pipe = StableDiffusionParadigmsPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5", scheduler=scheduler, torch_dtype=torch.float16
)
pipe = pipe.to("cuda")

ngpu, batch_per_device = torch.cuda.device_count(), 5
pipe.wrapped_unet = torch.nn.DataParallel(pipe.unet, device_ids=[d for d in range(ngpu)])

prompt = "a photo of an astronaut riding a horse on mars"
image = pipe(prompt, parallel=ngpu * batch_per_device, num_inference_steps=1000).images[0]
This pipeline improves sampling speed by running denoising steps in parallel, at the cost of increased total FLOPs. Therefore, it is better to call this pipeline when running on multiple GPUs. Otherwise, without enough GPU bandwidth sampling may be even slower than sequential sampling.

The two parameters to play with are parallel (batch size) and tolerance.

  • If it fits in memory, for 1000-step DDPM you can aim for a batch size of around 100 (e.g. 8 GPUs and batch_per_device=12 to get parallel=96). Higher batch size may not fit in memory, and lower batch size gives less parallelism.
  • For tolerance, using a higher tolerance may get better speedups but can risk sample quality degradation. If there is quality degradation with the default tolerance, then use a lower tolerance (e.g. 0.001).

For 1000-step DDPM on 8 A100 GPUs, you can expect around a 3x speedup by StableDiffusionParadigmsPipeline instead of StableDiffusionPipeline by setting parallel=80 and tolerance=0.1.

Diffusers also offers distributed inference support for generating multiple prompts in parallel on multiple GPUs. Check out the docs [here](https://huggingface.co/docs/diffusers/main/en/training/distributed_inference).

In contrast, this pipeline is designed for speeding up sampling of a single prompt (by using multiple GPUs).

StableDiffusionParadigmsPipeline

class diffusers.StableDiffusionParadigmsPipeline

< >

( vae: AutoencoderKL text_encoder: CLIPTextModel tokenizer: CLIPTokenizer unet: UNet2DConditionModel scheduler: KarrasDiffusionSchedulers safety_checker: StableDiffusionSafetyChecker feature_extractor: CLIPImageProcessor requires_safety_checker: bool = True )

Parameters

  • vae (AutoencoderKL) — Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
  • text_encoder (CLIPTextModel) — Frozen text-encoder. Stable Diffusion uses the text portion of CLIP, specifically the clip-vit-large-patch14 variant.
  • tokenizer (CLIPTokenizer) — Tokenizer of class CLIPTokenizer.
  • unet (UNet2DConditionModel) — Conditional U-Net architecture to denoise the encoded image latents.
  • scheduler (SchedulerMixin) — A scheduler to be used in combination with unet to denoise the encoded image latents. Can be one of DDIMScheduler, LMSDiscreteScheduler, or PNDMScheduler.
  • safety_checker (StableDiffusionSafetyChecker) — Classification module that estimates whether generated images could be considered offensive or harmful. Please, refer to the model card for details.
  • feature_extractor (CLIPImageProcessor) — Model that extracts features from generated images to be used as inputs for the safety_checker.

Parallelized version of StableDiffusionPipeline, based on the paper https://arxiv.org/abs/2305.16317 This pipeline parallelizes the denoising steps to generate a single image faster (more akin to model parallelism).

Pipeline for text-to-image generation using Stable Diffusion.

This model inherits from DiffusionPipeline. Check the superclass documentation for the generic methods the library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)

In addition the pipeline inherits the following loading methods:

as well as the following saving methods:

__call__

< >

( prompt: typing.Union[str, typing.List[str]] = None height: typing.Optional[int] = None width: typing.Optional[int] = None num_inference_steps: int = 50 parallel: int = 10 tolerance: float = 0.1 guidance_scale: float = 7.5 negative_prompt: typing.Union[typing.List[str], str, NoneType] = None num_images_per_prompt: typing.Optional[int] = 1 eta: float = 0.0 generator: typing.Union[torch._C.Generator, typing.List[torch._C.Generator], NoneType] = None latents: typing.Optional[torch.FloatTensor] = None prompt_embeds: typing.Optional[torch.FloatTensor] = None negative_prompt_embeds: typing.Optional[torch.FloatTensor] = None output_type: typing.Optional[str] = 'pil' return_dict: bool = True callback: typing.Union[typing.Callable[[int, int, torch.FloatTensor], NoneType], NoneType] = None callback_steps: int = 1 cross_attention_kwargs: typing.Union[typing.Dict[str, typing.Any], NoneType] = None debug: bool = False ) → StableDiffusionPipelineOutput or tuple

Parameters

  • prompt (str or List[str], optional) — The prompt or prompts to guide the image generation. If not defined, one has to pass prompt_embeds. instead.
  • height (int, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) — The height in pixels of the generated image.
  • width (int, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) — The width in pixels of the generated image.
  • num_inference_steps (int, optional, defaults to 50) — The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference.
  • parallel (int, optional, defaults to 10) — The batch size to use when doing parallel sampling. More parallelism may lead to faster inference but requires higher memory usage and also can require more total FLOPs.
  • tolerance (float, optional, defaults to 0.1) — The error tolerance for determining when to slide the batch window forward for parallel sampling. Lower tolerance usually leads to less/no degradation. Higher tolerance is faster but can risk degradation of sample quality. The tolerance is specified as a ratio of the scheduler’s noise magnitude.
  • guidance_scale (float, optional, defaults to 7.5) — Guidance scale as defined in Classifier-Free Diffusion Guidance. guidance_scale is defined as w of equation 2. of Imagen Paper. Guidance scale is enabled by setting guidance_scale > 1. Higher guidance scale encourages to generate images that are closely linked to the text prompt, usually at the expense of lower image quality.
  • negative_prompt (str or List[str], optional) — The prompt or prompts not to guide the image generation. If not defined, one has to pass negative_prompt_embeds instead. Ignored when not using guidance (i.e., ignored if guidance_scale is less than 1).
  • num_images_per_prompt (int, optional, defaults to 1) — The number of images to generate per prompt.
  • eta (float, optional, defaults to 0.0) — Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to schedulers.DDIMScheduler, will be ignored for others.
  • generator (torch.Generator or List[torch.Generator], optional) — One or a list of torch generator(s) to make generation deterministic.
  • latents (torch.FloatTensor, optional) — Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will ge generated by sampling using the supplied random generator.
  • prompt_embeds (torch.FloatTensor, optional) — Pre-generated text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, text embeddings will be generated from prompt input argument.
  • negative_prompt_embeds (torch.FloatTensor, optional) — Pre-generated negative text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, negative_prompt_embeds will be generated from negative_prompt input argument.
  • output_type (str, optional, defaults to "pil") — The output format of the generate image. Choose between PIL: PIL.Image.Image or np.array.
  • return_dict (bool, optional, defaults to True) — Whether or not to return a StableDiffusionPipelineOutput instead of a plain tuple.
  • callback (Callable, optional) — A function that will be called every callback_steps steps during inference. The function will be called with the following arguments: callback(step: int, timestep: int, latents: torch.FloatTensor).
  • callback_steps (int, optional, defaults to 1) — The frequency at which the callback function will be called. If not specified, the callback will be called at every step.
  • cross_attention_kwargs (dict, optional) — A kwargs dictionary that if specified is passed along to the AttentionProcessor as defined under self.processor in diffusers.cross_attention.
  • debug (bool, optional, defaults to False) — Whether or not to run in debug mode. In debug mode, torch.cumsum is evaluated using the CPU.

Returns

StableDiffusionPipelineOutput or tuple

StableDiffusionPipelineOutput if return_dict is True, otherwise a tuple. When returning a tuple, the first element is a list with the generated images, and the second element is a list of bools denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw) content, according to the safety_checker`.

Function invoked when calling the pipeline for generation.

Examples:

>>> import torch
>>> from diffusers import DDPMParallelScheduler
>>> from diffusers import StableDiffusionParadigmsPipeline

>>> scheduler = DDPMParallelScheduler.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="scheduler")

>>> pipe = StableDiffusionParadigmsPipeline.from_pretrained(
...     "runwayml/stable-diffusion-v1-5", scheduler=scheduler, torch_dtype=torch.float16
... )
>>> pipe = pipe.to("cuda")

>>> ngpu, batch_per_device = torch.cuda.device_count(), 5
>>> pipe.wrapped_unet = torch.nn.DataParallel(pipe.unet, device_ids=[d for d in range(ngpu)])

>>> prompt = "a photo of an astronaut riding a horse on mars"
>>> image = pipe(prompt, parallel=ngpu * batch_per_device, num_inference_steps=1000).images[0]

disable_vae_slicing

< >

( )

Disable sliced VAE decoding. If enable_vae_slicing was previously invoked, this method will go back to computing decoding in one step.

disable_vae_tiling

< >

( )

Disable tiled VAE decoding. If enable_vae_tiling was previously invoked, this method will go back to computing decoding in one step.

enable_model_cpu_offload

< >

( gpu_id = 0 )

Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared to enable_sequential_cpu_offload, this method moves one whole model at a time to the GPU when its forward method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with enable_sequential_cpu_offload, but performance is much better due to the iterative execution of the unet.

enable_sequential_cpu_offload

< >

( gpu_id = 0 )

Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, unet, text_encoder, vae and safety checker have their state dicts saved to CPU and then are moved to a torch.device('meta') and loaded to GPU only when their specific submodule has its forwardmethod called. Note that offloading happens on a submodule basis. Memory savings are higher than withenable_model_cpu_offload`, but performance is lower.

enable_vae_slicing

< >

( )

Enable sliced VAE decoding.

When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.

enable_vae_tiling

< >

( )

Enable tiled VAE decoding.

When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful to save a large amount of memory and to allow the processing of larger images.

←PaintByExample Pix2Pix Zero→