Search papers, labs, and topics across Lattice.
Shanghai Jiao Tong University
1
0
3
StreamKL slashes memory usage from quadratic to constant, enabling efficient long-context attention distillation on a single GPU.
