d5d232e207
This allows horovod to be built with frameworks=pytorch,tensorflow. I tracked down the crash I observed in #15719, where loading torch before tensorflow would cause a crash in: google::protobuf::internal::(anonymous namespace)::InitSCC_DFS(google::protobuf::internal::SCCInfoBase*) The solution is to make tensorflow compile against the protobuf version Spack provides, instead of allowing it to use it's own. It's likely we'll want to go after some of the others that are listed in third_party/systemlibs/syslibs_configure.bzl in the future. |
||
---|---|---|
.. | ||
spack |