PulseAugur
EN
LIVE 16:52:45

New research tackles FP4 pretraining stability with 2D block scaling

By PulseAugur Editorial · [1 sources] ·

A new research paper introduces a method for stable FP4 pretraining by addressing a critical issue with transpose-invariant 2D block scaling. Previous methods using 1D scaling groups suffered from scale inconsistency when matrices were transposed for backpropagation, leading to biased gradients. The proposed solution partitions matrices into square 2D blocks, ensuring that scale assignments remain consistent across forward and backward passes, thereby mitigating a specific source of systematic error in low-precision computations. AI

IMPACT Improves efficiency and stability in low-precision model training, potentially enabling larger models with reduced computational cost.

RANK_REASON The cluster contains a research paper detailing a novel technical approach to improving model pretraining. [lever_c_demoted from research: ic=1 ai=1.0]

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New research tackles FP4 pretraining stability with 2D block scaling

How we ranked this

Signal score
55 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
The cluster contains a research paper detailing a novel technical approach to improving model pretraining. [lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
paper, infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. dev.to — LLM tag TIER_1 English(EN) · Prabhakar Chaudhary ·

    Stable FP4 Pretraining with Transpose-Invariant 2D Block Scaling

    <h1> Stable FP4 Pretraining with Transpose-Invariant 2D Block Scaling </h1> <p><em>Billing Support — September 03, 2026</em></p> <p>A tensor value can receive one scale during the forward pass and a different scale after the same tensor is transposed for backpropagation. At FP4 p…

RELATED ENTITIES

RELATED TOPICS