yashagar-amd 2867294db3 fmha: fix coredump issue in small seqlen ...

PER_TOKEN_HEAD batch prefill reads q_descale_per_token for a full kM0
query tile. For small sequence lengths or chunked-prefill tail chunks,
some tile rows are padding rows beyond the valid query range, but the
kernel still indexed q_descale_per_token_ptr for those rows. In group
mode, q_descale_ptr also missed the per-batch query_start offset, so
later batches could read descales from the wrong position.

Fix the issue by offsetting q_descale_ptr with query_start, exposing
the mask's valid query row count, and guarding q_descale_per_token
loads for rows outside y_total.

Patch credits: Zhen Han <zhen.han@amd.com>

2026-06-05 11:05:38 +00:00

..

ck

[rocm-libraries] ROCm/rocm-libraries#7111 (commit 651947f)

2026-05-08 14:15:31 +00:00

ck_tile

fmha: fix coredump issue in small seqlen

2026-06-05 11:05:38 +00:00

rapidjson

Update pre-commit to fixed versions, run remod for ck_tile (#2895)

2025-10-16 15:29:17 -07:00