[KongchangAI]
Conceptpost-hoc compression

Post-hoc Compression

A technique that re-reduces the rank or reconstructs the parameters of a model or adapter after training to reduce storage. However, the reduced-rank structure may fail to translate into actual speedup due to padding or generic kernels.

Timeline (last 90 days)

Nov 3

After post-compression, a smaller rank does not automatically bring faster serving, because issues such as padding, factor access, and generic kernels can mask the retained low-rank structure.

Unverified50%

All Facts (1)

Source Articles