新研究通过二维块缩放解决 FP4 预训练稳定性问题

作者 PulseAugur 编辑部 · [1 个来源] · 2026-09-03 16:07

一篇新研究论文介绍了一种通过解决转置不变二维块缩放的关键问题来实现稳定 FP4 预训练的方法。先前使用一维缩放组的方法在矩阵转置进行反向传播时存在缩放不一致的问题，导致梯度偏差。提出的解决方案将矩阵划分为二维方形块，确保缩放分配在正向和反向传播中保持一致，从而减轻了低精度计算中系统性错误的特定来源。 AI

影响提高了低精度模型训练的效率和稳定性，有可能以更低的计算成本实现更大的模型。

排序理由该集群包含一篇详细介绍改进模型预训练的新技术方法的论文。[lever_c_demoted from research: ic=1 ai=1.0]

在 dev.to — LLM tag 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。我们如何撰写摘要 →

本文如何被排名

Signal score: 54 / 100; Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket: Tool; 该集群包含一篇详细介绍改进模型预训练的新技术方法的论文。[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.

完整方法见我们的编辑标准。

报道来源 [1]

dev.to — LLM tag TIER_1 English(EN) · Prabhakar Chaudhary · 2026-09-03 16:07

具有转置不变二维块缩放的稳定 FP4 预训练

<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…

本文如何被排名

报道来源 [1]

具有转置不变二维块缩放的稳定 FP4 预训练

相关实体

相关话题