Skip to content

fix: add head_dim=256 to fused SDPA full attention kernel #3293

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
Thump604 wants to merge 3 commits into
base: main
Choose a base branch
from
+13 −2
Open

fix: add head_dim=256 to fused SDPA full attention kernel #3293

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
726c9a0
fix: add head_dim=256 to fused SDPA full attention kernel
Thump604 Mar 22, 2026
7397435
perf: route head_dim=256 to unfused SDPA for short sequences
Thump604 Mar 24, 2026
d08d2da
feat: add head_dim=192 to fused SDPA kernel support
Thump604 Mar 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions mlx/backend/metal/kernels/steel/attn/kernels/steel_attention.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,6 +12,7 @@
attention, dtype, bq, bk, bd, wm, wn, mtype, float)

#define instantiate_attn_shapes_helper(iname, itype, mname, mtype) \
instantiate_attn(iname, itype, 32, 16, 256, 4, 1, mname, mtype) \
instantiate_attn(iname, itype, 32, 16, 128, 4, 1, mname, mtype) \
instantiate_attn(iname, itype, 32, 32, 80, 4, 1, mname, mtype) \
instantiate_attn(iname, itype, 32, 32, 64, 4, 1, mname, mtype)
Expand Down
3 changes: 2 additions & 1 deletion mlx/backend/metal/scaled_dot_product_attention.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -620,7 +620,8 @@ bool ScaledDotProductAttention::use_fallback(
(query_head_dim == 64 || query_head_dim == 96 || query_head_dim == 128 ||
query_head_dim == 256);
const bool sdpa_full_supported_head_dim = query_head_dim == value_head_dim &&
(query_head_dim == 64 || query_head_dim == 80 || query_head_dim == 128);
(query_head_dim == 64 || query_head_dim == 80 || query_head_dim == 128 ||
query_head_dim == 256);

const bool sdpa_full_supported_mask = !has_mask || has_arr_mask ||
(query_sequence_length <= key_sequence_length && do_causal);
Expand Down
Loading