Search papers, labs, and topics across Lattice.
1
0
3
Train massive MoEs on Hopper GPUs faster and with less memory, even without native FP4 support, by cleverly quantizing activations and communication.
