rules_cuda is a community effort for adding cuda support for bazel.
It currently supports:
linux and windows
nvcc+gcc or nvcc+msvc where cuda >= 10.0 or standalone clang cuda
relocatable device code (rdc) and device link time optimiaztion (dlto)
It is also purely implemented in starlark. Toolchains can be configured with cc_toolchain style DSL. Attacking TF repo might be too ambitious at the moment, I think OpenXLA might suit for it.
This is not yet feasible as XLA and Tensorflow are still tied together at Google, sharing the same code, but it might be possible in 1-2 months as they are more completely separated.
rules_cuda is a community effort for adding cuda support for bazel.
It currently supports:
It is also purely implemented in starlark. Toolchains can be configured with cc_toolchain style DSL. Attacking TF repo might be too ambitious at the moment, I think OpenXLA might suit for it.