← 返回事件
持续讨论AI发版

trunk/98059d9deac01fd3c485bdf60efac6b9121f0f5d: Fix oversized storage for reduced MSE loss (#199823)

图:PyTorch Releases

发生了什么

Linked issue or supporting maintainer Fixes #198951 BC-breaking? No. Summary (human written only) F.mse_loss with mean or sum reduction was returning a scalar that still held the full elementwise loss buffer in storage. This fixes the reduced-output setup so the result is allocated as a scalar from the start. smooth_l1_loss also has the same pattern, but I’m leaving that for a separate follow-up. Agent details (if a…

摘要按规则整理自下方来源原文

为什么在扩散

来源