Search papers, labs, and topics across Lattice.
University of Maryland
1
0
2
Predicting dynamic expert activations two layers in advance eliminates transfer stalls, enabling massive MoE models to run up to 11.7脳 faster on a single commodity GPU.
