Fix XNNPACK rejecting avg_pool2d with default arguments

[slipstr34m]

Problem

AvgPoolingConfig.check_constraints rejects every node with count_include_pad=True:

if count_include_pad:
    why(
        node,
        reason="zero-padding in the averaging calculation is not supported",
    )
    return False

But count_include_pad=True is the aten default, so a plain nn.AvgPool2d(2) — the most common way to write the layer — never delegates to XNNPACK. The delegate splits around it and the node falls back to the portable op, and the WhyNoPartition log tells the user their model has zero-padding it does not have:

WhyNoPartition: Node aten_avg_pool2d_default was not partitioned because zero-padding in the averaging calculation is not supported.

The flag only has semantics when padded elements exist: it controls whether zeros from padding count toward the averaging divisor. With the default padding=0 there is nothing to include or exclude, so the rejection excludes nodes XNNPACK handles correctly. Verified numerically: with padding=0, eager outputs for count_include_pad=True and False are bit-identical, and the delegated result matches eager within ordinary float tolerance; with padding=1 they genuinely differ, which is the case the backend does not implement.

Fix

Reject count_include_pad=True only when the node has non-zero padding, reading padding from the existing normalize_pool2d_args call:

kernel_size, _, padding, _ = normalize_pool2d_args(node, has_dilation=False)
...
if count_include_pad and any(p != 0 for p in padding):

A no-op for every node the config currently accepts (those all have count_include_pad=False), and every currently-rejected padded node is still rejected, so the only behavior change is crash-to-delegate for the default-argument case. The visitor needs no change: it serializes the padding values themselves and never reads count_include_pad, and XNNPACK's divisor convention is irrelevant when no padded elements exist.

Effect

Same model, nn.AvgPool2d(2) with all-default arguments, through to_edge_transform_and_lower with the XNNPACK partitioner:

before: delegates=0  remaining aten.avg_pool2d.default in graph, WhyNoPartition blames zero-padding
after:  delegates=1  no avg_pool2d left in graph, output matches eager
guard:  padding=(1,1) with count_include_pad=True still not delegated

Testing

New test_fp32_avgpool2d_default_args lowers nn.AvgPool2d(2) end to end through serialize and compares outputs against eager. Verified failing-first: with the tests kept and only the source reverted, exactly that test fails and the other seven pass, including test_fp32_avgpool2d_count_include_pad_unsupported (padding=(1,1)), which still rejects under the fix.

backends/xnnpack/test/ops/test_avgpool2d.py: 7 -> 8 passed. Full backends/xnnpack/test/ops/ suite: 229 passed on main, 230 passed with this change, no regressions. The suite runs exclude test_prelu.py, whose test_fp32_prelu_constant_weight_empty_decompositions_file_load segfaults identically on unmodified main in my environment (prebuilt runtime, macOS arm64), so it cannot observe this change either way. lintrunner reports no issues on both changed files.

cc @GregoryComer @digantdesai @cbilgin @JakeStevens

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21862

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit ddb9251 with merge base ed65b12 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[slipstr34m]

@pytorchbot label "release notes: none"

[github-actions]

This PR needs a release notes: label

If your change should be included in the release notes (i.e. would users of this library care about this change?), please use a label starting with release notes:. This helps us keep track and include your important work in the next release notes.

To add a label, you can comment to pytorchbot, for example
@pytorchbot label "release notes: none"

For more information, see
https://github.com/pytorch/pytorch/wiki/PyTorch-AutoLabel-Bot#why-categorize-for-release-notes-and-how-does-it-work.