test(sdpa): fuzz per-tensor ragged token-stride gaps in test_mhas_v2 - #516

Merged

vedaanta merged 4 commits into

NVIDIA:developfrom

vedaanta:vagarwalla/thd-kv-token-stride

Aug 11, 2026

Commits on Aug 11, 2026

test(sdpa): fuzz per-tensor ragged token-stride gaps in test_mhas_v2

vedaanta
and
claude
committed
8499013
test(sdpa): simplify ragged token gaps to whole tokens; always-on in the fp16 ragged sweeps

vedaanta
and
claude
committed
b7d8c10
test(sdpa): draw all four ragged token gaps up front; drop unused binding

vedaanta
and
claude
committed
17be556
test(sdpa): ragged token gaps on by default; auto-packed where not yet expressible

vedaanta
and
claude
committed
746ef17