avg_pool2d: count_include_pad reads node.args[5] but guards node.args[4] — IndexError on 5-arg nodes (ceil_mode=True), still present in 0.4.2

[xocialize]

Summary

replace_avg_pool2d in coreai_torch/_aten_to_core.py reads count_include_pad from
node.args[5] but guards on node.args[4]. A node carrying exactly five arguments passes
the guard and then raises IndexError: tuple index out of range, so conversion fails before any
tensor is produced.

Reproduced on coreai-torch 0.4.1 and 0.4.2 (latest), macOS 27.0 (26A5421a), M5 Max,
coreai-core==1.0.0b2.

Reproduction

import torch, torch.nn as nn
from coreai_torch import TorchConverter, get_decomp_table

class M(nn.Module):
    def __init__(self):
        super().__init__()
        self.p = nn.AvgPool2d(2, 2, 1, ceil_mode=True)   # ceil_mode=True is the trigger
    def forward(self, x):
        return self.p(x)

ep = torch.export.export(M().eval(), args=(torch.randn(1, 3, 32, 32),))
ep = ep.run_decompositions(dict(get_decomp_table()))
TorchConverter().add_exported_program(ep, input_names=["x"], output_names=["out"]).to_coreai()
# IndexError: tuple index out of range

torch.nn.functional.avg_pool2d(x, 2, 2, 1, True) fails identically.

Observed vs expected

Construction	node len(args)	0.4.1	0.4.2
nn.AvgPool2d(2, 2, 1, ceil_mode=True)	5	IndexError	IndexError
F.avg_pool2d(x, 2, 2, 1, True)	5	IndexError	IndexError
F.avg_pool2d(x, 2, 2, 1, True, False)	6	converts	converts
F.avg_pool2d(x, 2, 2, 0, False) (defaults)	3	converts	converts

Expected: all four convert. count_include_pad should default to True when absent.

Root cause

The signature is
aten.avg_pool2d(input, kernel_size, stride, padding, ceil_mode, count_include_pad, divisor_override),
so ceil_mode is index 4 and count_include_pad is index 5.

The two adjacent reads are:

ceil_mode = (
    node.args[4] if len(node.args) > 4 and node.args[4] is not None else False
)
count_include_pad = (
    node.args[5] if len(node.args) > 4 and node.args[4] is not None else True
)

The ceil_mode line is correct. The count_include_pad line reads index 5 while still
guarding index 4 — the guard was not updated when the line was copied. With exactly five
args the guard is satisfied (len > 4, args[4] is not None) and the read runs off the end.

Why it stays hidden

With a default ceil_mode, torch.export normalises the trailing arguments away and the node
carries only 3 args. The guard then fails cleanly and the else True branch happens to be
correct. The defect only surfaces when ceil_mode is explicitly non-default, which keeps element
4 in the graph while element 5 is still absent — so ordinary AvgPool2d usage converts fine and
ceil_mode=True does not.

Suggested fix

count_include_pad = (
    node.args[5] if len(node.args) > 5 and node.args[5] is not None else True
)

Workaround

Pass count_include_pad explicitly so the node carries six arguments:

F.avg_pool2d(x, kernel_size, stride, padding, ceil_mode, count_include_pad)

Note for downstream consumers

LibreYOLO carries a monkey-patch for this, scoped to
_AFFECTED_COREAI_TORCH_VERSIONS = {"0.4.1"}, which declines silently on any other version.
Since the defect is still present in 0.4.2, that shim now no-ops and the failure returns with no
diagnostic. Anyone pinning a version-scoped workaround for this will want to widen it until a fix
lands.