Skip to content

Add initial tuning for M5 pro and max #3211

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
angeloskath merged 2 commits into from
Mar 11, 2026
Merged

Add initial tuning for M5 pro and max #3211

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1efafed
Add gemm initial tuning for M5 pro and max
jagrit06 Mar 5, 2026
f775bca
Update tunings
jagrit06 Mar 10, 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
4 changes: 4 additions & 0 deletions mlx/backend/metal/kernels/steel/gemm/kernels/steel_gemm_fused_nax.metal
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,10 @@

#define instantiate_gemm_shapes_helper(iname, itype, oname, otype) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 64, 64, 256, 2, 2) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 64, 128, 64, 2, 4) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 64, 128, 256, 2, 4) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 128, 128, 64, 4, 4) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 128, 128, 256, 4, 4) \
instantiate_gemm_transpose_helper(iname, itype, oname, otype, 128, 128, 512, 4, 4)

instantiate_gemm_shapes_helper(float16, half, float16, half);
Expand Down
12 changes: 12 additions & 0 deletions mlx/backend/metal/matmul.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -204,6 +204,15 @@ void steel_matmul_regular_axpby_nax(
int bm = 128, bn = 128, bk = 512;
int wm = 4, wn = 4;

// Temp routing for larger devices
char devc = d.get_architecture().back();
if (devc == 's' || devc == 'c' || devc == 'd') {
bk = (K >= 8192 && K > (M + N)) ? 64 : 256;

bm = 64;
wm = 2;
}

// Prepare kernel name
std::ostringstream kname;

Expand Down Expand Up @@ -268,6 +277,9 @@ void steel_matmul_regular_axpby_nax(

// TODO: Explore device-based tuning for swizzle
int swizzle_log = tm <= 3 ? 0 : 1;
if (devc == 's' || devc == 'c' || devc == 'd') {
swizzle_log = 2;
}

// Prepare steel matmul params
GEMMParams params{/* const int M = */ M,
Expand Down
Loading