A training recipe combining Muon, single-scale RMSNorm, and a learnable embedding projection produces a 1.4B-parameter LLM with near-zero activation excess kurtosis and strong 4-bit quantization performance.
Croci, Marcelo Gennari do Nascimento, Torsten Hoefler, and James Hensman
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
-
Outlier-Safe Pre-Training for Robust 4-Bit Quantization of Large Language Models
A training recipe combining Muon, single-scale RMSNorm, and a learnable embedding projection produces a 1.4B-parameter LLM with near-zero activation excess kurtosis and strong 4-bit quantization performance.