Compiling PyTorch

Structure of a PyTorch build directory

A PyTorch build directory contains six subdirectories:

:- pytorch/

• torchbenchmark/
• torch-audio/
• torch-data/
• torch-text/
• torch-vision/

There is also the directory containing this file, torch-build/, which can be anywhere on your file system.

By default, the build directory is in ~/git but there are two ways to build PyTorch in other directories:

1. Setting the environment variable PYTORCH_BUILD_SUFFIX appends this value to the build directory, and also to the Conda environment which is used. For example, if PYTORCH_BUILD_SUFFIX=-grad, then the PyTorch build directory would be created in ~/git-grad and the Conda environment would be named pytorch-dev.

2. Or for finer grained control, you can independently set the environment variable PYTORCH_BUILD_DIRECTORY to set the build directory, PYTORCH_CONDA_ENV to set the name of the Conda environment

By default, PyTorch is cloned from [email protected]:pytorch/pytorch.git but you can override this by setting the environment variable PYTORCH_GIT_USER. For example, if PYTORCH_GIT_USER=octacat then the fork [email protected]:octacat/pytorch.git will be used.

Setting up the environment

• Set the correct CUDA version in pytorch-dev.yaml by changing the line cuda-version=12.2

• Create the conda environment: ./torch-env.sh

• [If you don't have them] Install the Nvidia drivers from https://www.nvidia.com/download/index.aspx

Python version. We set python=3.8 in pytorch-dev.yaml, as this is the minimum required version in PyTorch, and this disallows us from using features that are "too new". To debug some issues that may not reproduce on Python 3.8, you may need to create a different env with a newer Python version.

Building PyTorch and due diligence

• Have a read through the pytorch-* and torch-* scripts and edit them as needed.
  • You will at least need to set CUDA_PATH and TORCH_CUDA_ARCH_LIST correctly in torch-common.sh.
  • These scripts give you "sane defaults", but feel free to tailor them to your liking.
• Running torch-clone.sh will download PyTorch and all the domain libraries. If you just want PyTorch, you can edit the script accordingly.
• Running pytorch-build.sh will compile PyTorch.
• Running torch-build.sh will compile PyTorch, the domain libs, and torchbench.
• Running torch-update.sh checks out the last main in all the libraries. Useful if you haven't compiled in a while.

Running torchbench

Without making some of the following changes, benchmarks you run can be highly unstable, varying as much as 10% from run to run, even if you are running each benchmark multiple times. Note that you require root to be able to enact most of them.

GPU benchmarks

To run a torchbench model for CUDA devices on an A100 GPU, follow these steps:

0. Set export USE_FLASH_ATTENTION=1 and export USE_MEM_EFF_ATTENTION=1 in torch-common.py
1. Build pytorch and all the domain libraries with torch-build.sh (See above)
2. Lock the GPU clock rates by running sudo lock-clock-a100.sh
3. Launch the appropriate benchmark-runner with the relevant arguments, e.g.

PYTHONPATH=$HOME/git/torch-bench/ python benchmarks/dynamo/torchbench.py \
  --performance --inductor --train --amp --only hf_GPT2

In the same directory there are also huggingface.py and timm_models.py which are run in a similar manner.

CPU benchmarks

If using an AWS instance (g4dn.metal), there is a script used by the Meta team for their benchmarks which is found in the torchbench repo. You can run it with the command

sudo $(which python) torchbenchmark/util/machine_config.py --configure

For other machines, a similar result can be achieved manually by following these steps:

1. Disable hyperthreading. Look at what the set_hyper_threading function in the torchbenchmark/util/machine_config.py does.
2. Disable Turbo Boost. The CPU might not have it, if the directory /sys/devices/system/cpu/intel_pstate does not exist, no need to do anything. If it does exist, look at set_intel_no_turbo_state and set_pstate_frequency in machine_config.py.
3. Set Intel c-state to 1. You need to edit /etc/default/grub and add intel_idle.max_cstate=1 to the GRUB_CMDLINE_LINUX_DEFAULT variable. Then run sudo update-grub and reboot.
4. CPU core isolation. This might not be strictly necessary if you can make sure there are no other processes running in the machine when running the benchmarks. The idea is to tell the OS not use some CPU cores at all unless they are specifically requested by taskset. Note that if you do this it will make all other workflows (such as compilation) slower since they will have less cores they can use. To do this follow the same steps as in previous point but instead of intel_idle.max_cstate=1 add isolcpus=6-11 where 6-11 is the range of cores you want to isolate.