---
id: 20260730-T0-07
title: "稳定FP4训练：通过换位不变块量化突破LLM低精度训练瓶颈"
title_en: "Stable FP4 Training for LLMs via Transposition-Invariant Block Quantization"
url: https://ai.daily.yangsir.net/daily/20260730-T0-07
issue_date: 2026-07-30
publish_date: 2026-07-29T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2607.24953
---

# 稳定FP4训练：通过换位不变块量化突破LLM低精度训练瓶颈

arXiv 论文解决4位浮点（FP4）LLM训练不稳定问题。研究发现，现有的块量化方法因对张量矩阵的转置敏感，导致训练过程中梯度波动剧烈。作者提出一种换位不变块量化方案，在保持FP4精度的同时，显著提升训练稳定性，使低精度训练能够成功收敛。

## English Version

**Stable FP4 Training for LLMs via Transposition-Invariant Block Quantization**

An arXiv paper addresses the instability issue in FP4 training for LLMs. It identifies that existing block quantization methods are sensitive to tensor transpose, causing gradient instability. The authors propose a transposition-invariant block quantization scheme that maintains FP4 precision while significantly improving training stability, enabling successful convergence.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2607.24953)

**详情页**：https://ai.daily.yangsir.net/daily/20260730-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*