Skip to content

Commit e60f241

Browse files
authored
metal : FA support F32 K and V and head size = 32 (ggml-org#16531)
* metal : FA support F32 K and V and head size = 32 * graph : remove obsolete comment [no ci]
1 parent e38b7c6 commit e60f241

File tree

4 files changed

+112
-58
lines changed

4 files changed

+112
-58
lines changed

ggml/src/ggml-metal/ggml-metal-device.m

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -693,7 +693,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
693693
return true;
694694
case GGML_OP_FLASH_ATTN_EXT:
695695
// for new head sizes, add checks here
696-
if (op->src[0]->ne[0] != 40 &&
696+
if (op->src[0]->ne[0] != 32 &&
697+
op->src[0]->ne[0] != 40 &&
697698
op->src[0]->ne[0] != 64 &&
698699
op->src[0]->ne[0] != 80 &&
699700
op->src[0]->ne[0] != 96 &&

0 commit comments

Comments
 (0)