Ck tile/paged attention workaround (#1894)

* Correction in GetRangeAlongX()

* Work-around to solve the failures in test_paged_attention_ck in xformers
This commit is contained in:
Qianfeng
2025-02-17 14:29:25 +08:00
committed by GitHub
parent 6b6fcd370b
commit a3757a5f9c
2 changed files with 2 additions and 2 deletions

View File

@@ -310,7 +310,7 @@ struct SimplifiedGenericAttentionMask
const index_t x_per_split = ck_tile::max(1, integer_divide_ceil(x_total, num_splits));
const index_t split_start = x_per_split * i_split;
const index_t split_end = split_start + x_per_split;
const index_t split_end = ck_tile::min(x_total, split_start + x_per_split);
return ck_tile::make_tuple(ck_tile::max(origin_start, split_start),
ck_tile::min(origin_end, split_end));