Complete chunkwise GatedDeltaNet#91

Open

learning-chip wants to merge 73 commits into

linear_attnfrom

Commits on Apr 15, 2026

static chunk GDN source from tilelang
learning-chip
committed
standalone static chunk GDN
learning-chip
committed
chain all kernels together to test e2e GDN
learning-chip
committed
use unified PTO_LIB_PATH
learning-chip
committed
BSND varlen version of chunk_cumsum
learning-chip
committed
partial porting of dynamic chunk_h, wy_fast, kkt
learning-chip
committed
partial working dynamic chunk_h
learning-chip
committed
finish chunk_o part
learning-chip
committed
fix scaled_dot_kkt functionality without hybriding torch hlpers
learning-chip
committed
merge kkt into one kernel launch
learning-chip
committed
checkpointing todo items and lessons
learning-chip
committed
attempt to debug wy_fast
learning-chip
committed
wy fast now works correctly
learning-chip
committed
finish chunk_h and update notes
learning-chip
committed
add skill template for general NPU kernel dev

jiawei_zhuang
authored and
learning-chip
committed
fix typo in skill

jiawei_zhuang
authored and
learning-chip
committed
rewrite Mandatory requirements

jiawei_zhuang
authored and
learning-chip
committed
mark highly recommended practices

jiawei_zhuang
authored and
learning-chip
committed
performance measurement of static tilelang reference
learning-chip
committed
update static_baseline shape and benchmark result
learning-chip
committed
update triton reference benchmark numbers
learning-chip
committed
rename dynamic bsnd dir
learning-chip
committed

Commits on Apr 16, 2026

Commits on Apr 17, 2026

Commits on Apr 19, 2026

Commits on Apr 20, 2026

Commits on Apr 25, 2026

avoid expensive sync and stream query inside kernel call
learning-chip
committed

Commits on Apr 28, 2026