Add single-dispatch layer-by-layer multi-head attention - #91

Draft

andrej wants to merge 8 commits into

andrej:mha-lxl-sd

Commits on Jul 24, 2026

Add single-dispatch layer-by-layer multi-head attention (mha_prefill_lxl)
andrej
committed
46b5aef
Fix CI benchmark marker + address PR #91 review comments
andrej
committed
efa6239
fix sync regression
andrej
committed
f967fd5
make test reporting more resilient
andrej
committed
a9ccf22
Address PR #91 review: revert unrelated sequence.py changes, trim softmax_stats comment
andrej
committed
493859b
Fix reinterleave StridedCopy dim/stride mismatch causing device-DMA heap corruption
andrej
committed
5134cb9
mha_prefill_lxl: read outputs via dispatch-agnostic read_host()
andrej
committed
6ca9659
softmax: restrict heavy single-core partial-softmax test to NPU2
andrej
committed
f8774a0