Skip to content

Fix Metal row reductions on negative-stride views #4267

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 2 commits into from
Aug 16, 2026
+12 −2
Merged

Fix Metal row reductions on negative-stride views #4267

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
f06ca82
Fix Metal row reductions on negative-stride views
FU-max-boop Aug 15, 2026
2bf38e3
Run negative-stride reduction test on all backends
FU-max-boop Aug 16, 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
3 changes: 2 additions & 1 deletion mlx/backend/metal/kernels/reduction/reduce_row.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -337,7 +337,8 @@ template <

// lid.x * N_READS breaks the per_thread_row_reduce interface a bit. Maybe it
// needs a small refactor.
in += elem_to_loc<IdxT>(out_idx, shape, strides, ndim) + lid.x * N_READS;
in +=
elem_to_loc<IdxT>(out_idx, shape, strides, ndim) + IdxT(lid.x) * N_READS;

LoopedElemToLoc<NDIMS, IdxT, (NDIMS > 2)> loop(reduce_ndim);
const device T* row;
Expand Down
11 changes: 10 additions & 1 deletion python/tests/test_reduce.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,5 @@
# Copyright © 2023 Apple Inc.

import unittest
from itertools import combinations, permutations

import mlx.core as mx
Expand Down Expand Up @@ -47,6 +46,16 @@ def test_expand_sums(self):
np.allclose(z_npy, np.array(z_mlx), atol=1e-4)
)

def test_row_reduce_negative_stride(self):
x_npy = np.arange(1, 131).reshape(2, 65)[::-1]
x_mlx = mx.arange(1, 131).reshape(2, 65)[::-1]

for op in ["sum", "max", "min", "mean", "var"]:
with self.subTest(op=op):
expected = getattr(np, op)(x_npy, axis=-1)
actual = getattr(mx, op)(x_mlx, axis=-1)
self.assertTrue(np.allclose(expected, actual))

def test_dtypes(self):
int_dtypes = [
"int8",
Expand Down
Loading