Using a rounded Gaussian noise distribution for pseudo-quantization makes low-precision FP weight training stable and cheap, matching or approaching BF16 baseline loss in LLM pretraining.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Gaussian Weight Sampling for Scalable, Efficient and Stable Pseudo-Quantization Training
Using a rounded Gaussian noise distribution for pseudo-quantization makes low-precision FP weight training stable and cheap, matching or approaching BF16 baseline loss in LLM pretraining.