Skip to content

Add a fused full-attention path for head_dim 256 on NAX devices #3842

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

Merged
zcbenz merged 6 commits into from
Aug 19, 2026
Merged

Add a fused full-attention path for head_dim 256 on NAX devices #3842

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
55ea7b6
Add a fused full-attention path for head_dim 256 on NAX devices
wyanzhao Jul 13, 2026
a8be3b3
Split the NAX head-dim-split attention path into its own kernel
wyanzhao Aug 15, 2026
667bb61
Drop the head_dim 256 padding copies; read KV-cache slices past their…
wyanzhao Aug 16, 2026
7d9ce26
nit
zcbenz Aug 18, 2026
5193924
nit
zcbenz Aug 19, 2026
b86d328
Merge branch 'main' into nax-256-dsplit
wyanzhao Aug 19, 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
11 changes: 10 additions & 1 deletion benchmarks/python/sdpa_bench.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -215,9 +215,18 @@ def get_gflop_count(B, M, N, K):
( 1, 4096, 5000, 128, 32, 8),
( 1, 2048, 32121, 128, 32, 8),
)

shapes_256 = (
# ( B, qsl, ksl, head_dim, n_qh, n_kvh)
( 1, 1024, 1024, 256, 24, 4),
( 1, 2048, 2048, 256, 24, 4),
( 1, 4096, 4096, 256, 24, 4),
( 1, 4096, 5000, 256, 24, 4),
( 1, 2048, 32121, 256, 24, 4),
)
# fmt: on

shapes = shapes_64 + shapes_72 + shapes_80 + shapes_96 + shapes_128
shapes = shapes_64 + shapes_72 + shapes_80 + shapes_96 + shapes_128 + shapes_256

masks = [None, "bool", "causal"]

Expand Down
5 changes: 3 additions & 2 deletions mlx/backend/metal/jit_kernels.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1330,7 +1330,8 @@ MTL::ComputePipelineState* get_steel_attention_nax_kernel(
int bd,
int wm,
int wn,
const array& m) {
const array& m,
bool split_d) {
const auto& lib_name = kernel_name;
auto lib = d.get_library(lib_name, [&]() {
std::string kernel_source;
Expand All @@ -1340,7 +1341,7 @@ MTL::ComputePipelineState* get_steel_attention_nax_kernel(
metal::steel_attention_nax(),
get_template_definition(
lib_name,
"attention_nax",
split_d ? "attention_nax_dsplit" : "attention_nax",
get_type_string(q.dtype()),
bq,
bk,
Expand Down
3 changes: 2 additions & 1 deletion mlx/backend/metal/kernels.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -426,7 +426,8 @@ MTL::ComputePipelineState* get_steel_attention_nax_kernel(
int bd,
int wm,
int wn,
const array& m);
const array& m,
bool split_d);

// Create a GPU kernel template definition for JIT compilation
template <typename... Args>
Expand Down
Loading