metal : FA support F32 K and V and head size = 32 (#16531)

* metal : FA support F32 K and V and head size = 32

* graph : remove obsolete comment [no ci]
This commit is contained in:
Georgi Gerganov authored and GitHub committed 2025-10-13 23:07:57 +03:00
1 parent e38b7c6e9e
commit e60f241eac
4 files changed
+106 -52

No files matched your search

-1
View File
@@ -1323,7 +1323,6 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_tensor * cur;
// TODO: replace hardcoded padding with ggml-provided padding
if (cparams.flash_attn && kq_b == nullptr) {
GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet");