For the foreseeable future (until distributions are migrated), we would like travis CI to build off of PyTorch master, rather than the release branch.

Steps for a potential strategy:

• Use travis cache to cache the packages installed via pip.
• Build PyTorch's master branch in travis and cache this build to be used for all later travis builds.

Edit - the strategy above did not work for multiple reasons - travis's cache is partitioned by environment keys (bloats the number of binaries stored), suffers from potential cache corruption issues, does not play well with staged builds, takes too long to build PyTorch etc. In the end, we decided to build the binaries separately and upload them to S3.