Build PyTorch from source

(You may want to see the already built releases before building yourself)

This project aims to build PyTorch from the source using a Dockerfile. The idea is to simplify the build so that one can choose the the Cuda/CuDNN version which better fits ones environment.

As PyTorch pre-built binaries require specific CUDA/CuDNN versions, e.g., #10971 you can't have pre-built PyTorch with CUDA 9.1 after the 0.4.0 release. For instance if you are using Debian Stretch and the Nvidia distribution packages you won't be able to use newer versions as they are built with CUDA 9.2.

The default configuration of the Dockerfile will build:

• Python 3.5
• PyTorch 1.3
• Cuda 9.1.85
• CuDNN 7.1.3

As it downloads the packages straight from Nvidia you can build the one you prefer with Cuda, the package comes straight from here. The following versions are then available, just replace the CUDA_VERSION variable by one of the following:

• 8.0.44
• 8.0.61
• 9.0.176
• 9.1.85
• 9.2.88
• 9.2.148
• 10.0.130
• 10.1.105
• 10.1.168
• 10.1.243

Likewise, CuDNN version can be one of the following. Just replace the CUDNN_PKG_VERSION variable to the one you prefer. The packages come directly from here.

• 7.0.1.13-1+cuda8.0
• 7.0.2.38-1+cuda8.0
• 7.0.3.11-1+cuda8.0
• 7.0.3.11-1+cuda9.0
• 7.0.4.31-1+cuda8.0
• 7.0.4.31-1+cuda9.0
• 7.0.5.15-1+cuda8.0
• 7.0.5.15-1+cuda9.0
• 7.0.5.15-1+cuda9.1
• 7.1.1.5-1+cuda8.0
• 7.1.1.5-1+cuda9.0
• 7.1.1.5-1+cuda9.1
• 7.1.2.21-1+cuda8.0
• 7.1.2.21-1+cuda9.0
• 7.1.2.21-1+cuda9.1
• 7.1.3.16-1+cuda8.0
• 7.1.3.16-1+cuda9.0
• 7.1.3.16-1+cuda9.1
• 7.1.4.18-1+cuda8.0
• 7.1.4.18-1+cuda9.0
• 7.1.4.18-1+cuda9.2
• 7.2.1.38-1+cuda8.0
• 7.2.1.38-1+cuda9.0
• 7.2.1.38-1+cuda9.2
• 7.3.0.29-1+cuda9.0
• 7.3.0.29-1+cuda10.0
• 7.3.1.20-1+cuda9.0
• 7.3.1.20-1+cuda9.2
• 7.3.1.20-1+cuda10.0
• 7.4.1.5-1+cuda9.0
• 7.4.1.5-1+cuda9.2
• 7.4.1.5-1+cuda10.0
• 7.4.2.24-1+cuda9.0
• 7.4.2.24-1+cuda9.2
• 7.4.2.24-1+cuda10.0
• 7.5.0.56-1+cuda9.0
• 7.5.0.56-1+cuda9.2
• 7.5.0.56-1+cuda10.0
• 7.5.0.56-1+cuda10.1
• 7.5.1.10-1+cuda9.0
• 7.5.1.10-1+cuda9.2
• 7.5.1.10-1+cuda10.0
• 7.5.1.10-1+cuda10.1
• 7.6.0.64-1+cuda9.0
• 7.6.0.64-1+cuda9.2
• 7.6.0.64-1+cuda10.0
• 7.6.0.64-1+cuda10.1
• 7.6.1.34-1+cuda9.0
• 7.6.1.34-1+cuda9.2
• 7.6.1.34-1+cuda10.0
• 7.6.1.34-1+cuda10.1
• 7.6.2.24-1+cuda9.0
• 7.6.2.24-1+cuda9.2
• 7.6.2.24-1+cuda10.0
• 7.6.2.24-1+cuda10.1
• 7.6.3.30-1+cuda9.0
• 7.6.3.30-1+cuda9.2
• 7.6.3.30-1+cuda10.0
• 7.6.3.30-1+cuda10.1
• 7.6.4.38-1+cuda9.0
• 7.6.4.38-1+cuda9.2
• 7.6.4.38-1+cuda10.0
• 7.6.4.38-1+cuda10.1

One thing to note is that some of the packages will respect the Cuda version you are looking for, e.g.,

apt-get install ...
    cuda-command-line-tools-9-1 \
    cuda-cublas-dev-9-1 \
    cuda-cudart-dev-9-1 \
    cuda-cufft-dev-9-1 \
    cuda-curand-dev-9-1 \
    cuda-cusolver-dev-9-1 \
    cuda-cusparse-dev-9-1 \

have the 9-1 in the package names, if you are building a Cuda 10.1 then you should replace the 9-1 everywhere by 10-1.

Building

Just run (sudo) docker build -t build-torch . in the same directory as the one where the Dockerfile is stored. The pytorch wheel will be stored on the /pytorch/dist directory of the container.

Getting the wheel out of the container

There are several ways to copy data from container to the host machine. Here is a suggestion:

sudo docker cp CONTAINER_ID:/tmp/pip/torch-1.3.0a0+50c90a2-cp35-cp35m-linux_x86_64.whl .

To get the CONTAINER_ID just run docker ps -alq or docker ps then see the CONTAINER_ID respective to the build-torch image, or the one you -t image tag you gave in the building command.