Add manywheel2014-builder images

[mattip]

This PR

• adds a Dockerfile_2014 that uses the pre-build manylinux2014 image as a base, meaning no need to build python, auditwheel, and patchelf
• tries to remain true to the previous docker image so that building pytorch should JustWork
• convert the build script into a loop for both docker image bases, which should create two sets of docker images

If this looks OK and actually uploads some images, I think the next iterative steps would be a PR to use the images (tagged with manylinux2014-builder:XXX) in pytorch/pytorch, and then to come back and fix the image failures.

[mattip]

A first step toward #670 which goes further toward enabling #520

[seemethere]

I think if we're going to do this it should just be an env variable that you pass through instead of iterating a for loop here, something like:

MANYLINUX_VERSION=2014 manywheel/build_docker.sh

Just so that we don't have to build all of the variants every single time.

This logic to build all of the variants can be added to manywheel/build_all_docker.sh

[mattip]

adopted

[mattip]

A maintainer needs to approve the CI run

[mattip]

ping @seemethere

[mattip]

I updated the BASE_CUDA_VERSION to 10.2.

Should the update of cuda 10.1 to 10.2 also be done in the other places it appears:

libtorch/ubuntu16.04/Dockerfile
manywheel/Dockerfile
manywheel/build.sh

[mattip]

@malfet @seemethere any thing else needed?

[snnn]

Not after long you will need to move to manylinux2 because the latest CUDA release requires libstdc++ from GCC 5.x but the one you got in manylinux2014 is based on GCC 4.x.

"CUDA Math Libraries toolchain uses C++11 features, and a C++11-compatible standard library (libstdc++ >= 20150422) is required on the host." from: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html

[snnn]

As you are using manylinux2014_x86_64 as the base, it should already have cmake and gcc. So do you need to install it again?

[mattip]

If it is already installed, yum will not install it again.

[snnn]

manylinux installs cmake from pypi. yum doesn't know that.
And, they don't have gcc7, they installed gcc 9. So when you install GCC7, you will get two GCCs.

[mattip]

Not after long you will need to move to manylinux2_24 ...

... which apparently also will not be enough, see the discussion in issue #520. But this will be a first step in that direction.

[mattip]

Any ideas what happened to github actions?

[rgommers]

Any ideas what happened to github actions?

That's not very useful CI output:

Never seen that before. Just restart it by close/open or pushing a new commit?

[mattip]

@seemethere @malfet could you start CI?

[mattip]

It crashed again. I rebased off master maybe that will solve it.
@seemethere @malfet please restart when you get a chance