Add cudnn.fla: a cuDNN drop-in for flash-linear-attention (GDN + KDA) - #596

Merged

YangXu1990uiuc merged 8 commits into

NVIDIA:developfrom

YangXu1990uiuc:yanxu/fla-compat

Aug 17, 2026

Commits on Aug 17, 2026

Add cudnn.fla: a cuDNN-accelerated drop-in for flash-linear-attention GDN

YangXu1990uiuc
and
claude
committed
6cb8f0b
Add KDA (Kimi Delta Attention) to cudnn.fla

YangXu1990uiuc
and
claude
committed
d0baa29
Add an end-to-end hybrid-model perf-share / support-gap benchmark

YangXu1990uiuc
and
claude
committed
21dc3af
Address CodeRabbit review

YangXu1990uiuc
and
claude
committed
8cf0a6b
cudnn.fla: fuse L2-norm/gate/beta in-kernel via the #616 native flags

YangXu1990uiuc
and
claude
committed
8ce4cda
cudnn.fla: decline all non-bf16 KDA inputs, not just fp16

YangXu1990uiuc
and
claude
committed
5dc3f00
Drop the e2e perf-share benchmark (moved to PR #609)

YangXu1990uiuc
and
claude
committed
3c0c194
cudnn.fla: lazily export `fla` from the top-level package

YangXu1990uiuc
and
claude
committed
a4211b9