Search papers, labs, and topics across Lattice.
1
0
3
6
Training LLMs at exascale is now more practical: custom GPU kernels and a sharded optimizer push MoE training speed by 1.7x.
