Building Linux wheels with CUDA 11.0
The build scripts in this repo add the -Xcompiler -fno-gnu-unique flags
as suggested in pytorch/pytorch#52663 .
Docker is required to build the linux wheels. The wheels can be build as follows
# go to the builder root folder and start a docker container with the
# pytorch/manywheel-cuda110 image
sudo docker run -it --ipc=host --rm -v $(pwd):/remote pytorch/manylinux-cuda110 bash
# inside the container go to /remote/manywheel
cd /remote/manywheel
# define target python and pytorch version
export DESIRED_PYTHON=3.7
export PYTORCH_BUILD_VERSION=1.7.1
./build.sh
# after this the wheel is located in the wheelhouse110 directory in the
# repository root.original README.md contents:
pytorch builder
Scripts to build pytorch binaries and do end-to-end integration tests.
Folders:
- conda : files to build conda packages of pytorch, torchvision and other dependencies and repos
- manywheel : scripts to build linux wheels
- wheel : scripts to build OSX wheels
- windows : scripts to build Windows wheels
- cron : scripts to drive all of the above scripts across multiple configurations together
- analytics : scripts to pull wheel download count from our AWS s3 logs
- test_community_repos : scripts that test built binaries with various downstream repos that are high-priority and use pytorch
- test_imports_docker : some smoke tests to import torch in combination with other package imports, to check binary stability / compatibility