P-Cast Precision in FP8 Attention: Sink-Induced Collapse and the Optimality of S=2^8 文章

ArXiv CS.AI2026-06-08NEWSen作者: Reed Lau

详细信息

来源站点
ArXiv CS.AI
作者
Reed Lau
文章类型
NEWS
语言
en
发布日期
2026-06-08

摘要

arXiv:2606.06521v1 Announce Type: cross Abstract: FP8 (E4M3) acceleration for attention computation offers significant throughput gains, but the 3-bit mantissa introduces precision challenges when the softmax probability matrix P is cast to FP8 before the P*V matrix multiplication. We analyze two implementation choices that affect output precision under the Attention Sink phenomenon: (1) the KV block iteration order, and (2) the static scaling factor applied to P before casting. We show that forward KV iteration causes "P-collapse" -- to leading order, a fraction Phi(Delta + delta_k - 6.93 - ln S) of non-sink P values underflow to zero, where the small shift delta_k ~ 1 (for k_sink = 4) is the expected within-sink-block score maximum -- and that reverse iteration removes it, with a zero-underflow guarantee when reverse is combined with S = 256.

相关事件

暂无数据

相关公司

暂无数据

相关人物

暂无数据

相关产品

暂无数据