Fix BatchNorm fusion producing invalid ONNX when Conv nodes share weight initializers

[Copilot]

When a model reuses Conv2d+BatchNorm2d blocks (same weights called twice), the BatchNorm fusion rewrite creates a new initializer with the same name as the shared weight, overwriting it in the graph's initializer dict. This sets _is_initializer = False on the original value still referenced by the second Conv node, producing an invalid model:

InvalidArgument: Node input 'conv.weight_1' is not a graph input, initializer, or output of a previous node.

Reproducer:

class MWE(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv = nn.Conv2d(1, 16, kernel_size=3, padding=1)
        self.bn = nn.BatchNorm2d(16)

    def forward(self, x):
        f_1 = self.bn(self.conv(x[:, 0:1]))
        f_2 = self.bn(self.conv(x[:, 1:2]))
        return f_1 + f_2

torch.onnx.export(MWE().eval(), ..., optimize=True)  # Invalid model

Changes

• _fuse_batchnorm.py: Added a check in _FuseBatchNormBase.check() that rejects the fusion when the inbound node's weight/bias initializers are used by nodes outside the matched pattern. This prevents the overwrite of shared initializers.
• _fuse_batchnorm_test.py: Added regression test with two Conv+BN pairs sharing the same weight and bias initializers.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 72.06%. Comparing base (1ef0ec9) to head (5c5277f).
⚠️ Report is 2 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2883      +/-   ##
==========================================
+ Coverage   72.04%   72.06%   +0.01%     
==========================================
  Files         239      239              
  Lines       29305    29324      +19     
  Branches     2880     2884       +4     
==========================================
+ Hits        21112    21131      +19     
  Misses       7216     7216              
  Partials      977      977              

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[gramalingam]

Looks ok. I think it would help to have a utility to do this kind of destructive transformation (the rule shouldn't have done this in the first place). Specifically, an utility could implement replace_uses_of_initializer_in( old_initializer_value, new_value, nodes) ... to abstract this logic. If the initializer has no uses outside the set of nodes, we can simply alter/swap. Otherwise, we can create a new initializer with the new value and a new (unique) name and use that instead. This sort of improvement can be done in a later PR though.