Emergency split CUDA libtorch build/test into separate job (#26859)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26859
CUDA builds are intermittently taking greater than five hours,
hitting CircleCI's timeout limit, and also all around making
developers unhappy. Part of the reason for this is because
they build PyTorch twice: once as normal, and once as libtorch.
This diff splits libtorch into a new job to parallelize this
and get us below the patch. It's an emergency diff because
I did the minimum possible work to make this work, including
grody hacks to make sure macos libtorch builds still work
(without adding a separate job there).
- Add a new libtorch config, to cuda9 (same as before). Disable
generation of the other test variants.
- Adjust common.sh to NO LONGER set BUILD_TEST_LIBTORCH for
pytorch-linux-trusty-py3.6-gcc7; we will test for *libtorch*
in the job name for this case. (I noticed a bug while
looking at this.)
- Adjust build.sh and test.sh. The eventual logic is that if you are a
*libtorch* build, ONLY build libtorch; otherwise do the same
thing you used to do (including respecting BUILD_TEST_LIBTORCH)
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Test Plan: Imported from OSS
Differential Revision: D17810592
Pulled By: ezyang
fbshipit-source-id: 8dcdb8f7424ddda293500d9fc90097a54dca28b9
