Search papers, labs, and topics across Lattice.
1
0
2
MatrixFSDP slashes optimizer-step latency by over 54x while enabling training on models that exceed 80 GB GPU limits.
