I need to use a more recent version of Pytorch with CUDA 10.2. Release 20.03 is the most recent one with CUDA 10.02 support, however its version of pytorch (1.5.0a0+8f84ded) does not match the pytorch 1.5.0 release. I would like to update the image with a higher version of pytorch (1.5.0 or 1.5.1). Within the image, the example docker file for a patch is below. However, there are no directions on how to write the my-pytorch-modification.patch file. What should be included here to appropriately change the pytorch version?
When I run the image and attempt to update pytorch with (conda install pytorch=1.5.0 -c pytorch) it fails with a huge number of incompatibilities.
# # This example Dockerfile illustrates a method to apply # patches to the source code in NVIDIA's PyTorch # container image and to rebuild PyTorch. The RUN command # included below will rebuild PyTorch in the same way as # it was built in the original image. # # By applying customizations through a Dockerfile and # `docker build` in this manner rather than modifying the # container interactively, it will be straightforward to # apply the same changes to later versions of the PyTorch # container image. # # https://docs.docker.com/engine/reference/builder/ # FROM nvcr.io/nvidia/pytorch:20.03-py3 # Bring in changes from outside container to /tmp # (assumes my-pytorch-modifications.patch is in same directory as Dockerfile) COPY my-pytorch-modifications.patch /tmp # Change working directory to PyTorch source path WORKDIR /opt/pytorch # Apply modifications RUN patch -p1 < /tmp/my-pytorch-modifications.patch # Rebuild PyTorch RUN cd pytorch && \ TORCH_CUDA_ARCH_LIST="5.2 6.0 6.1 7.0 7.5+PTX" \ CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" \ NCCL_INCLUDE_DIR="/usr/include/" \ NCCL_LIB_DIR="/usr/lib/" \ python setup.py install && python setup.py clean # Reset default working directory WORKDIR /workspace