[KongchangAI]
Concept

DAMP

A quantization scheme for the recurrent state of linear attention models that rounds the state immediately after each generated token. It supports eight-bit precision, but exhibits cumulative error problems at four-bit precision.

Timeline (last 90 days)

Oct 22

PACE with mixed 4-bit and 8-bit heads performs comparably to 8-bit DAMP on three larger models while using only about two-thirds of its memory.

Unverified50%

All Facts (1)

Source Articles