MLCommons Algorithmic Efficiency is a benchmark and competition measuring neural network training speedups due to algorithmic improvements in both training algorithms and models.
From meeting minutes from Michael Shi: Challenge is ensuring that JAX and PyTorch are equivalent. PyTorch should be doable by changing the DDP wrapper to the FSDP wrapper.
It is useful to shard optimizer state across devices (to save significant memory). This reflects current practice. We want to support it.