Search papers, labs, and topics across Lattice.
1
0
3
HGA enables fine-tuning on sequences of 16,384 tokens with just 15.28 GB of VRAM, outperforming traditional methods constrained by memory limits.