Testing infrastructure for ROCm Communication Collectives Library (RCCL).
The RCCL test suite provides following frameworks along with the existing rccl-UnitTests TestBed framework:
Following is a new testing framework for running single node & single process test in isolation:
Run tests in isolated processes with clean environment settings.
Base class for multi-process distributed tests using MPI. Logging: environment-driven per-rank log files (RCCL_MPI_LOG_ALL_RANKS), TEST_* macros with NCCL_DEBUG, and scoped MPIHelpers::TestLogAssertionContext for asserting NCCL lines (see the summary tables in the doc).