【第633期】Lighthouse Attention重构长文本AI底层逻辑 episode artwork

EPISODE · Jun 24, 2026 · 15 MIN

【第633期】Lighthouse Attention重构长文本AI底层逻辑

from Seventy3

Seventy3:借助NotebookLM的能力进行论文解读,专注人工智能、大模型、机器人算法、crypto方向,让大家跟着AI一起进步。如果你想要解读自己的论文,获得更多曝光度。请联系小助手微信:seventy3_podcast 加群。合作邮箱:zhiwudazhanjiangshi#gmail.com今天的主题是:Long Context Pre-Training with Lighthouse AttentionSummary在极长序列长度下训练因果 Transformer(causal transformers)受到缩放点积注意力(SDPA)平方级时间和内存复杂度的瓶颈制约。在这项工作中,我们提出了 Lighthouse Attention,这是一种仅在训练阶段使用的、基于对称选择的分层注意力算法。它包裹在普通的 SDPA 周围,并且可以在训练接近尾声时轻松移除。我们的分层选择还是无梯度的,这使我们无需处理复杂且可能效率低下的反向传播算子(backward pass kernel)。我们的贡献主要有三个方面:一个亚平方(subquadratic)级的分层预处理和后处理步骤,用于对序列进行自适应压缩和解压缩。一种对称压缩策略,在保持自左向右因果性的同时,同步对 query、key 和 value 进行池化(pooling),这大大提高了并行度。一种两阶段训练方法:在绝大部分时间内使用 Lighthouse Attention 进行预训练,最后通过简短的训练恢复为全注意力(full attention)模型。我们进行了初步的小规模大语言模型(LLM)预训练实验,结果表明,在所有其他设置均匹配的情况下,与全注意力训练相比,我们的方法非常有效,不仅实现了更快的总训练时间,而且在恢复阶段后获得了更低的最终损失(loss)。原文链接:https://arxiv.org/abs/2605.06554前往小宇宙评论区与主播互动

Episode metadata supplied by the publisher feed · Published Jun 24, 2026

Embed this episode

NOW PLAYING

【第633期】Lighthouse Attention重构长文本AI底层逻辑

0:00 15:23

No transcript for this episode yet

We transcribe on demand. Request one and we'll notify you when it's ready — usually under 10 minutes.

No similar episodes found.

No similar podcasts found.

Frequently Asked Questions

How long is this episode of Seventy3?

This episode is 15 minutes long.

When was this Seventy3 episode published?

This episode was published on June 24, 2026.

Can I download this Seventy3 episode?

Yes. Use the download control on the episode player to save the publisher-provided media file.
URL copied to clipboard!