Search papers, labs, and topics across Lattice.
1
0
2
Sub-quadratic linear attention no longer requires full-precision state caching: 8-bit recurrent states can match FP32 accuracy while delivering up to 3.70× kernel speedups without any retraining.
