PyTorch Build for Apple Silicon

NOTE: This document is useful only for scratch build from source code for Apple Silicon.

Work-flow

• install prerequisites
• download PyTorch source from repository
• install PyTorch through a wheel-build, or standard install

Test Environment

• Python 3.9.8
• macOS Montorey (ver.12.0.1)
• PyTorch 1.12 (2022/02/23@JST)

Prerequisite Installation

- CMake

NOTE: Compilation needs version 3.13 or greater of cmake.

$brew install cmake

- OpenBLAS

$brew install openblas

- Yaml

$python3 -m pip install pyyaml

Download the source-code of PyTorch

git clone --recursive https://github.com/pytorch/pytorch

Enter the directory;

$cd pytorch

Setting CMakeList.txt

pThead is not supported on OS-X;

set(THREADS_PREFER_PTHREAD_FLAG OFF)

OS-X does not support Microsoft VisualStudio, so comment-out between these lines;

if(MSVC)

and

endif(MSVC)

And, setting APPLE as TRUE.

At line 184, Apple Silicon does not support NVIDIA's CUDA;

option(USE_CUDA "Use CUDA" OFF)

Apple Silicon does not support NVIDIA's cuDNN;

USE_CUDNN "Use cuDNN" OFF

At line 193, Apple Silicon does not support AMD's ROCm;

option(USE_ROCM "Use ROCm" OFF)

At line 230, Apple Silicon does not support NVIDIA's NCCL Library;

    USE_NCCL "Use NCCL" OFF

At line 232, Apple Silicon does not support AMD's RCCL Library;

cmake_dependent_option(USE_RCCL "Use RCCL" OFF

At line 243, Apple Silicon does not support NUMA Memory;

    USE_NUMA "Use NUMA. Only available on Linux." OFF

At line 252, Apple Silicon does not support OpenMP;

option(USE_OPENMP "Use OpenMP for parallel code" OFF)

Build for Wheel-based Install

- Compile Sorce-Code

Compile with below command;

$python3 setup.py build

- Build Wheel

$python3 setup.py bdist bdist_wheel

- Install

Wheel file is in "dist" directory. You can use pip install for the wheel file.

Build for Standard-Install

- Compile Sorce-Code

Compile with below command;

$python3 setup.py build

- Install

python3 setup.py develop && python3 -c "import torch"