Search papers, labs, and topics across Lattice.
1
0
2
Forget massive MoE parameter counts: ButterflyMoE slashes memory requirements by 150x using learned rotations of a shared, quantized prototype, without sacrificing accuracy.
