Skip to content

Arm backend: Fix avg_pool2d divisor_override with ceil_mode and no padding #24894

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
zingo merged 2 commits into from
Nov 15, 2026
+15 −2
Merged

Arm backend: Fix avg_pool2d divisor_override with ceil_mode and no padding #24894

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
e6c7ba9
Arm backend: Fix avg_pool2d divisor_override with ceil_mode and no pa…
EylonKrause Nov 11, 2026
7fb7fb7
Merge branch 'main' into fix/avgpool-divisor-override-ceil-mode
zingo Nov 15, 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
9 changes: 7 additions & 2 deletions backends/arm/_passes/decompose_avg_pool2d_pass.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,12 @@ def _compute_post_pad(
divisor_override,
) -> int | torch.SymInt:

if pad == 0:
# Without padding, count_include_pad=True divides a window that runs past
# the end of the input by its valid elements only, as the decomposed
# count_include_pad=False avg_pool2d does. A divisor_override needs full
# windows though, so the part of a ceil_mode window outside the input must
# still be padded with zeros.
if pad == 0 and not (ceil_mode and divisor_override is not None):
return pad
pad_adjust = adjust_pooling_pad_if_needed(size, kernel, stride, pad, ceil_mode)

Expand Down Expand Up @@ -141,7 +146,7 @@ def call_operator(self, op, args, kwargs, meta):
divisor_override,
)

if count_include_pad and (pad_h > 0 or pad_w > 0):
if count_include_pad and any(p != 0 for p in pad):
if op in aten_avg_pool2d:
pad_args = (x, pad, "constant", 0.0)
else:
Expand Down
8 changes: 8 additions & 0 deletions backends/arm/test/ops/test_avg_pool2d.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,14 @@ def forward(self, x: torch.Tensor):
AvgPool2d(3, 2, 1, True, True, divisor_override=2),
(torch.rand(1, 1, 14, 14),),
),
"divisor_override_ceil_mode_no_padding": lambda: (
AvgPool2d(3, 2, 0, True, False, divisor_override=2),
(torch.rand(1, 1, 15, 14),),
),
"divisor_override_ceil_mode_no_padding_count_include_pad": lambda: (
AvgPool2d(3, 2, 0, True, True, divisor_override=4),
(torch.rand(1, 1, 14, 14),),
),
"non_divisible_no_padding": lambda: (
AvgPool2d(3, 2, 0),
(torch.rand(1, 16, 56, 56),),
Expand Down
Loading