Lion Cub compresses Lion updates with L1 quantization and sparse momentum synchronization, reducing distributed training time by up to 5.1x at similar convergence.
We trained a 735M parameter GPT model on OpenWebText with 16 GPUs using a cluster equipped with 4 H100 GPUs per node
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2024 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Lion Cub: Minimizing Communication Overhead in Distributed Lion
Lion Cub compresses Lion updates with L1 quantization and sparse momentum synchronization, reducing distributed training time by up to 5.1x at similar convergence.