Search papers, labs, and topics across Lattice.
1
0
2
Stochastic rounding excels in Transformer MLPs, while round-to-nearest is better at the output head, enabling a 28% perplexity reduction via mixed-precision emulation.
