trends.barnas.net

tool · first seen 4 Oct 2026

fused kernel for loss and gradient

Also: Pallas kernels

Custom accelerator kernel that computes loss and gradients on-chip without materializing intermediate tensors in high-bandwidth memory.

Mentions per week

0119 May11 Aug10 NovPapersCodeDiscussion

PapersCodeDiscussion

Earliest sources

Latest sources

Mentioned together with

sequential recommendationbinary cross-entropy losschunked loss computation