Signals 4 · free daily AI digest

UE5M3 FP4 Block Scaling for Stable Language Model Pretraining

Paper recorded by Signals 4 on 2026-09-02 in cs.LG. Abstract reproduced from arXiv; link to the original below.

Published 2026-09-02 on arXiv · recorded by Signals 4 on 2026-09-03

Category: cs.LG · 机器学习 · first seen 2026-09-03

Abstract

Stable 4-bit floating-point (FP4) pretraining is difficult because the E2M1 payload represents only a narrow range of magnitudes. NVIDIA's Transformer Engine \nv{} recipe addresses this with current-tensor scaling, a randomized Hadamard transform (RHT), and bfloat16 (BF16) final layers, adding work outside the FP4 matrix multiplications. We instead pair E2M1 payloads with unsigned E5M3 (\ue{}) blo

Read on arXiv →

#147 most recent of 215 cs.LG papers we have recorded · ↑ newer: The Implications of Linguistic Illegibility for LLM Security · ↓ older: Learning Spectral-Like Mesh-Free Discretisations
Cite this page: UE5M3 FP4 Block Scaling for Stable Language Model Pretraining: the #147 most recent of 215 cs.LG papers we have recorded (as of 2026-09-02). Source: Signals 4 (Signals API) — https://data.jiangzhang.ca/signals4/t/papers/ue5m3-fp4-block-scaling-for-stable-language-model-pretraining.html
Free to quote with attribution to “Signals 4 (Signals API)”. Machine-readable: papers.json
Related: More cs.LG papers · arXiv signals · All papers · Today in AI
Get 4 AI signals a day by email — free.
Subscribe free → See all plans →
Get 4 AI signals a day by email — free
All models · All repos · By company · Daily editions