ReleaseNVIDIANVIDIApublished Oct 30, 2026seen 9h

NVIDIA/cudnn-frontend v1.32.1

NVIDIA/cudnn-frontend

Open original ↗

Captured source

source ↗
GH
GitHub/github.com/NVIDIA/cudnn-frontend
NVIDIA/cudnn-frontend v1.32.1
Source ↗
published Oct 30, 2026seen 9hcaptured 9hhttp 200method plain

v1.32.1-release

Repository: NVIDIA/cudnn-frontend

Tag: v1.32.1

Published: 2026-10-30T06:45:56Z

Prerelease: no

Release notes:

cuDNN Frontend v1.32.1 Release Notes

This patch release expands DSA backward support and improves CSA backward performance and memory use.

  • DSA backward: Add native H96 support for Top-K widths 2049–2051 on SM100/SM103 in both ordinary and deterministic modes (#2042).
  • DSA indexer backward: Add native H32 support to the opt-in sm100_v2 backend on SM100/SM103, including unaligned Top-K widths in ordinary mode (#1908).
  • CSA backward: Speed up packed ratio-2/4 metadata lookup, especially for batches with many fragmented sequences (#1932).
  • CSA workspace: Reduce deterministic ratio-4 backward workspace by about two-thirds for long contexts with at least 65,536 compressed rows (#1968).