Skip to content

Latest commit

 

History

History
206 lines (160 loc) · 7.38 KB

070_INSTALL_XFORMERS.md

File metadata and controls

206 lines (160 loc) · 7.38 KB
Raw
title
Installing xFormers

:material-image-size-select-large: Installing xformers

xFormers is toolbox that integrates with the pyTorch and CUDA libraries to provide accelerated performance and reduced memory consumption for applications using the transformers machine learning architecture. After installing xFormers, InvokeAI users who have CUDA GPUs will see a noticeable decrease in GPU memory consumption and an increase in speed.

xFormers can be installed into a working InvokeAI installation without any code changes or other updates. This document explains how to install xFormers.

Pip Install

For both Windows and Linux, you can install xformers in just a couple of steps from the command line.

If you are used to launching invoke.sh or invoke.bat to start InvokeAI, then run the launcher and select the "developer's console" to get to the command line. If you run invoke.py directly from the command line, then just be sure to activate it's virtual environment.

Then run the following three commands:

pip install xformers==0.0.16rc425
pip install triton
python -m xformers.info output

The first command installs xformers, the second installs the triton training accelerator, and the third prints out the xformers installation status. If all goes well, you'll see a report like the following:

xFormers 0.0.16rc425
memory_efficient_attention.cutlassF:               available
memory_efficient_attention.cutlassB:               available
memory_efficient_attention.flshattF:               available
memory_efficient_attention.flshattB:               available
memory_efficient_attention.smallkF:                available
memory_efficient_attention.smallkB:                available
memory_efficient_attention.tritonflashattF:        available
memory_efficient_attention.tritonflashattB:        available
swiglu.fused.p.cpp:                                available
is_triton_available:                               True
is_functorch_available:                            False
pytorch.version:                                   1.13.1+cu117
pytorch.cuda:                                      available
gpu.compute_capability:                            8.6
gpu.name:                                          NVIDIA RTX A2000 12GB
build.info:                                        available
build.cuda_version:                                1107
build.python_version:                              3.10.9
build.torch_version:                               1.13.1+cu117
build.env.TORCH_CUDA_ARCH_LIST:                    5.0+PTX 6.0 6.1 7.0 7.5 8.0 8.6
build.env.XFORMERS_BUILD_TYPE:                     Release
build.env.XFORMERS_ENABLE_DEBUG_ASSERTIONS:        None
build.env.NVCC_FLAGS:                              None
build.env.XFORMERS_PACKAGE_FROM:                   wheel-v0.0.16rc425
source.privacy:                                    open source

Source Builds

xformers is currently under active development and at some point you may wish to build it from sourcce to get the latest features and bugfixes.

Source Build on Linux

Note that xFormers only works with true NVIDIA GPUs and will not work properly with the ROCm driver for AMD acceleration.

xFormers is not currently available as a pip binary wheel and must be installed from source. These instructions were written for a system running Ubuntu 22.04, but other Linux distributions should be able to adapt this recipe.

1. Install CUDA Toolkit 11.7

You will need the CUDA developer's toolkit in order to compile and install xFormers. Do not try to install Ubuntu's nvidia-cuda-toolkit package. It is out of date and will cause conflicts among the NVIDIA driver and binaries. Instead install the CUDA Toolkit package provided by NVIDIA itself. Go to CUDA Toolkit 11.7 Downloads and use the target selection wizard to choose your platform and Linux distribution. Select an installer type of "runfile (local)" at the last step.

This will provide you with a recipe for downloading and running a install shell script that will install the toolkit and drivers. For example, the install script recipe for Ubuntu 22.04 running on a x86_64 system is:

wget https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda_11.7.0_515.43.04_linux.run
sudo sh cuda_11.7.0_515.43.04_linux.run

Rather than cut-and-paste this example, We recommend that you walk through the toolkit wizard in order to get the most up to date installer for your system.

2. Confirm/Install pyTorch 1.13 with CUDA 11.7 support

If you are using InvokeAI 2.3 or higher, these will already be installed. If not, you can check whether you have the needed libraries using a quick command. Activate the invokeai virtual environment, either by entering the "developer's console", or manually with a command similar to source ~/invokeai/.venv/bin/activate (depending on where your invokeai directory is.

Then run the command:

python -c 'exec("import torch\nprint(torch.__version__)")'

If it prints 1.13.1+cu117 you're good. If not, you can install the most up to date libraries with this command:

pip install --upgrade --force-reinstall torch torchvision

3. Install the triton module

This module isn't necessary for xFormers image inference optimization, but avoids a startup warning.

pip install triton

4. Install source code build prerequisites

To build xFormers from source, you will need the build-essentials package. If you don't have it installed already, run:

sudo apt install build-essential

5. Build xFormers

There is no pip wheel package for xFormers at this time (January 2023). Although there is a conda package, InvokeAI no longer officially supports conda installations and you're on your own if you wish to try this route.

Following the recipe provided at the xFormers GitHub page, and with the InvokeAI virtual environment active (see step 1) run the following commands:

pip install ninja
export TORCH_CUDA_ARCH_LIST="6.0;6.1;6.2;7.0;7.2;7.5;8.0;8.6"
pip install -v -U git+https://github.com/facebookresearch/xformers.git@main#egg=xformers

The TORCH_CUDA_ARCH_LIST is a list of GPU architectures to compile xFormer support for. You can speed up compilation by selecting the architecture specific for your system. You'll find the list of GPUs and their architectures at NVIDIA's GPU Compute Capability table.

If the compile and install completes successfully, you can check that xFormers is installed with this command:

python -m xformers.info

If suiccessful, the top of the listing should indicate "available" for each of the memory_efficient_attention modules, as shown here:

memory_efficient_attention.cutlassF:               available
memory_efficient_attention.cutlassB:               available
memory_efficient_attention.flshattF:               available
memory_efficient_attention.flshattB:               available
memory_efficient_attention.smallkF:                available
memory_efficient_attention.smallkB:                available
memory_efficient_attention.tritonflashattF:        available
memory_efficient_attention.tritonflashattB:        available
[...]

You can now launch InvokeAI and enjoy the benefits of xFormers.

Windows

To come

(c) Copyright 2023 Lincoln Stein and the InvokeAI Development Team