Aluminum provides a generic interface to high-performance communication libraries with a focus on allreduce algorithms. Blocking and non-blocking algorithms and GPU-aware algorithms are supported. Aluminum also contains custom implementations of select algorithms to optimize for certain situations. The v1.1.0 release includes improvements in benchmarking/testing infrastructure, better progress engine binding on HIP/ROCm systems, a project logo, and more. v1.2.0 adds better support for low-precision data.
Learn more:
