GPAS scales down intermediate activations while preserving backward gradients, reducing activation variance growth in Pre-LN transformers and improving pretraining convergence and downstream performance.
Outlier weighed layerwise sparsity (owl): A missing secret sauce for pruning llms to high sparsity
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
support 1representative citing papers
citing papers explorer
-
GPAS: Accelerating Convergence of LLM Pretraining via Gradient-Preserving Activation Scaling
GPAS scales down intermediate activations while preserving backward gradients, reducing activation variance growth in Pre-LN transformers and improving pretraining convergence and downstream performance.