Building pytorch using conda

build.sh

#!/bin/bash

# Activate the build environment
eval "$(conda shell.bash hook)"
conda activate pytorch-dev

cd ~/git/pytorch

# Enable ccache
export CCACHE_COMPRESS=true
export CMAKE_C_COMPILER_LAUNCHER=ccache
export CMAKE_CXX_COMPILER_LAUNCHER=ccache
export CMAKE_CUDA_COMPILER_LAUNCHER=ccache

# Configure the pytorch build
export CMAKE_PREFIX_PATH=$CONDA_PREFIX
export USE_CUDA=1
export USE_XNNPACK=1
export USE_DISTRIBUTED=0
export USE_NCCL=0
export MAX_JOBS=24
export CUDA_HOME=/usr/local/cuda-10.2.89
export DEBUG=1

# Cleanup any old torch installs
pip uninstall torch -y

# Build pytorch and install to the curren environment (pytorch-dev)
python setup.py install

pytorch-dev.yaml

name: pytorch-dev 
channels: 
  - conda-forge 
  - pytorch 
dependencies: 
  - python=3.8 
  - numpy
  - ninja 
  - pyyaml 
  - mkl 
  - mkl-include 
  - setuptools 
  - cmake 
  - cffi 
  - typing 
  - pytest 
  - magma-cuda102 
  - pytest 
  - hypothesis 
  - ccache 
  - flake8

setup_commands.md

Setting up a development environment to build and test pytorch

1. Install miniconda into your home directory:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
chmod +x ./Miniconda3-latest-Linux-x86_64.sh
./Miniconda3-latest-Linux-x86_64.sh

Then follow the setup prompts. You may need to restart your terminal after installing to update your environment variables.

2. Setup conda environment and clone pytorch:

conda env create -f pytorch-dev.yaml
git clone https://github.com/pytorch/pytorch
git -C pytorch submodule update --init --recursive

3. Build using the included script:

./build.sh

After this finishes, pytorch will be installed in the conda environment pytorch-dev which can be enabled with

conda activate pytorch-dev