You don't need any special flags to build the tfrt_cc_binary targets under //backends/gpu/....
Those contain only host code for GPUs, and the same binary dynamically loads and calls both ROCm and CUDA libraries.
Only the cuda_library targets contain GPU device code and won't build for AMD GPUs. The tfrt_cc_binary targets under //backends/gpu/... don't depend on these, but tools like bef_executor do if you bazel build with --//:enable_gpu. If you don't have the CUDA SDK installed on your machine, leave //:enable_gpu disabled (which is the default).
FullZing
commented
1 year ago
You don't need any special flags to build the tfrt_cc_binary targets under
//backends/gpu/.... Those contain only host code for GPUs, and the same binary dynamically loads and calls both ROCm and CUDA libraries.Only the
cuda_librarytargets contain GPU device code and won't build for AMD GPUs. Thetfrt_cc_binarytargets under//backends/gpu/...don't depend on these, but tools likebef_executordo if youbazel buildwith--//:enable_gpu. If you don't have the CUDA SDK installed on your machine, leave//:enable_gpudisabled (which is the default).I hope this helps.