single mention · optimization
Unified pluggable activation storage manager
A unified storage abstraction in which every tensor saved for the backward pass is associated with a pluggable storage backend.1 Recomputation, quantization, and offload become composable storage policies declared via lightweight annotations on tensors, decoupled from the model code.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
every tensor saved for the backward pass is associated with a pluggable storage backend
— k3 tech reportRecomputation, quantization, and offload/remote- offload are merely storage policies under this abstraction and can be freely composed at tensor granularity; policies are declared via lightweight annotations on tensors, fully decoupled from the model code
— k3 tech report
Filed under optimization › training runtime