LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
LeapQuant, a paper from researchers at UC Berkeley, the University of Washington, MIT, Perplexity AI, and NVIDIA, proposes a training-free way to quantize the recurrent state of linear attention layers to 8 bits.