Stable Fast

Introduction

NOTE: stable-fast is only in beta stage and is prone to be buggy, feel free to try it out and give suggestions!

What is this?

stable-fast is an ultra lightweight inference optimization library for HuggingFace Diffusers on NVIDIA GPUs. stable-fast provides super fast inference optimization by utilizing some key techniques and features:

• CUDNN Convolution Fusion: stable-fast implements a series of fully-functional and fully-compatible CUDNN convolution fusion operators for all kinds of combinations of Conv + Bias + Add + Act computation patterns.
• Low Precision & Fused GEMM: stable-fast implements a series of fused GEMM operators that compute with fp16 precision, which is fast than PyTorch's defaults (read & write with fp16 while compute with fp32).
• NHWC & Fused GroupNorm: stable-fast implements a highly optimized fused NHWC GroupNorm + GELU operator with OpenAI's triton, which eliminates the need of memory format permutation operators.
• Fully Traced Model: stable-fast improves the torch.jit.trace interface to make it more proper for tracing complex models. Nearly every part of StableDiffusionPipeline can be traced and converted to TorchScript. It is more stable than torch.compile and has a significantly lower CPU overhead than torch.compile and supports ControlNet and LoRA.
• CUDA Graph: stable-fast can capture the UNet structure into CUDA Graph format, which can reduce the CPU overhead when the batch size is small.
• Fused Multihead Attention: stable-fast just uses xformers and make it compatible with TorchScript.

Differences With Other Acceleration Libraries

• Fast: stable-fast is specialy optimized for HuggingFace Diffusers. It achieves a high performance across many libraries.
• Minimal: stable-fast works as a plugin framework for PyTorch. it utilizes existing PyTorch functionality and infrastructures and is compatible with other acceleration techniques, as well as popular fine-tuning techniques and deployment solutions.

Optimize StableDiffusionPipeline

import torch
from diffusers import (StableDiffusionPipeline, EulerAncestralDiscreteScheduler)
from sfast.compilers.stable_diffusion_pipeline_compiler import (compile,
                                                                CompilationConfig
                                                                )

def load_model():
    model = StableDiffusionPipeline.from_pretrained(
        'runwayml/stable-diffusion-v1-5', torch_dtype=torch.float16)
    model.scheduler = EulerAncestralDiscreteScheduler.from_config(
        model.scheduler.config)
    model.safety_checker = None
    model.to(torch.device('cuda'))
    return model

model = load_model()

config = CompilationConfig.Default()

# xformers and triton are suggested for achieving best performance.
# It might be slow for triton to generate, compile and fine-tune kernels.
try:
    import xformers
    config.enable_xformers = True
except ImportError:
    print('xformers not installed, skip')
try:
    import triton
    config.enable_triton = True
except ImportError:
    print('triton not installed, skip')
# CUDA Graph is suggested for small batch sizes.
# After capturing, the model only accepts one fixed image size.
# If you want the model to be dynamic, don't enable it.
config.enable_cuda_graph = True

compiled_model = compile(model, config)

kwarg_inputs = dict(
    prompt=
    '(masterpiece:1,2), best quality, masterpiece, best detail face, lineart, monochrome, a beautiful girl',
    height=512,
    width=512,
    num_inference_steps=30,
    num_images_per_prompt=1,
)

# NOTE: Warm it up.
# The first call will trigger compilation and might be very slow.
# After the first call, it should be very fast.
output_image = compiled_model(**kwarg_inputs).images[0]

# Let's see the second call!
output_image = compiled_model(**kwarg_inputs).images[0]

Performance Comparison

Performance varies very greatly across different hardware/software/platform/driver configurations. It is very hard to benchmark accurately. And preparing the environment for benchmarking is also a hard job. I have tested on some platforms before but the results may still be inaccurate.

RTX 4090 (512x512, batch size 1, fp16, tcmalloc enabled)

(??): OneFlow seems to be not working well with SD 2.1

RTX 3080 Ti (512x512, batch size 1, fp16, tcmalloc enabled)

(??): OneFlow seems to be not working well with SD 2.1

RTX 3090 (512x512, batch size 1, fp16, tcmalloc enabled)

A100

Sorry, currently A100 is hard and expensive to rent from cloud server providers in my region.

Benchmark results will be available when I have the access to A100 again.

Compatibility

Usage

Installation

NOTE: stable-fast is currently only tested on Linux. You need to install PyTorch with CUDA support at first (versions from 1.12 to 2.1 are suggested).

Install From Source

# Make sure you have CUDNN/CUBLAS installed.
# https://developer.nvidia.com/cudnn
# https://developer.nvidia.com/cublas

# Install PyTorch with CUDA and other packages at first
pip3 install 'torch>=1.12.0' 'diffusers>=0.19.3' 'xformers>=0.0.20' 'triton>=2.1.0'

# (Optional) Makes the build much faster
pip3 install ninja

# Set TORCH_CUDA_ARCH_LIST if running and building on different GPU types
pip3 install -v -U git+https://github.com/chengzeyi/stable-fast.git@main#egg=stable-fast
# (this can take dozens of minutes)

NOTE: Any usage outside sfast.compilers is not guaranteed to be backward compatible. NOTE: To get the best performance, xformers and OpenAI's triton>=2.1.0 need to be installed and enabled. You might need to build xformers from source to make it compatible with your PyTorch.

Some Common Methods To Speed Up PyTorch

# TCMalloc is highly suggested to reduce CPU overhead
# https://github.com/google/tcmalloc
LD_PRELOAD=/path/to/libtcmalloc.so python3 ...

import packaging.version
import torch

if packaging.version.parse(torch.__version__) >= packaging.version.parse('1.12.0'):
    torch.backends.cuda.matmul.allow_tf32 = True