Looks super cool. Does it remove the need for manual gradient checkpointing? Also curious if there are expected to be memory / speed improvements if you're usi

Looks super cool. Does it remove the need for manual gradient checkpointing? Also curious if there are expected to be memory / speed improvements if you're using it on a single GPU or if most gains come from improved parallelism across devices.