perf(moe): warp-split moe_q4_grouped_mma variant, opt-in (BUTTER_LAGUNA_MOE_NSPLIT=1) - #75

Merged

TheTom merged 9 commits into

devfrom

perf/moe-nsplit-occupancy

Jul 29, 2026

Commits on Jul 25, 2026

fix(sdpa): per-block K-prep for sliding-window prefill attention
TheTom
committed
f7c0ab2
perf(sdpa): shrink windowed-layer KV block size 2048->1024
TheTom
committed
0425f8b
perf(prefill): size-adaptive default chunk (min(n,2048))
TheTom
committed
f0106ab
perf(marlin): pad smem_WT row stride 128->136 to cut bank conflicts
TheTom
committed
bb548cc
perf(moe): O(n log n) heapsort in moe_build_tiles, replacing O(n^2) selection sort
TheTom
committed
b2ffe9d
perf(moe): deepen moe_q4_grouped_mma cp.async pipeline 2->3 stages
TheTom
committed
2ffc2b6
feat(ops): wire moe_gather_q4/_swiglu + rms_norm_f16out/gated_group_rmsnorm_batched_f16out to iron@dev
TheTom
committed
48e0672
chore(deps): bump iron lock to dev#131b199 (PR #76 pool-cap 4->8 GiB fix)
TheTom
committed
5afc3a1

Commits on Jul 26, 2026

perf(moe): warp-count-doubled moe_q4_grouped_mma_nsplit (BUTTER_LAGUNA_MOE_NSPLIT=1)
TheTom
committed
f34145c