Fix ShapeInferencePass crash on initializers without const_value

[Copilot]

ShapeInferencePass fails with AssertionError when any graph initializer has const_value=None — the standard state for models built without loaded weights. This blocks shape inference on any pre-weight-loading model.

W = ir.Value(name="W", shape=ir.Shape([3, 4]), type=float_type)  # no const_value
graph = ir.Graph(..., initializers=[W], ...)
common_passes.ShapeInferencePass()(model)  # AssertionError in _c_api_utils.py:46

Changes

• _c_api_utils.py: Replace assert initializer.const_value is not None with a conditional — skip shape/dtype inference from const_value when absent, remove the initializer from the dict (avoiding serialization failure), but keep it as a graph input so ONNX shape inference can still use its declared type/shape.
• _c_api_utils.py (restore path): Use initializers.add() instead of register_initializer() for unloaded initializers, since register_initializer() enforces const_value is not None.
• shape_inference_test.py: Add test with a MatMul node where one input is an initializer with no const_value, verifying inference succeeds and model state is properly restored.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.05%. Comparing base (5fd5e94) to head (3de5306).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #380      +/-   ##
==========================================
+ Coverage   80.02%   80.05%   +0.03%     
==========================================
  Files          52       52              
  Lines        6394     6399       +5     
  Branches     1294     1297       +3     
==========================================
+ Hits         5117     5123       +6     
  Misses        912      912              
+ Partials      365      364       -1     

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

[Copilot]

Pull request overview

Fixes a crash in ShapeInferencePass when a graph contains “unloaded” initializers (const_value=None), which is common for models constructed before weights are loaded. The update ensures ONNX shape inference can still run using the initializer’s declared type/shape.

Changes:

• Update _c_api_utils.call_onnx_api() to tolerate initializers with const_value=None by skipping const-derived shape/dtype inference and temporarily treating them as graph inputs for the ONNX C API call.
• Adjust restoration logic to re-add unloaded initializers without using register_initializer() (which requires const_value).
• Add a regression test covering MatMul shape inference with an unloaded initializer.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
src/onnx_ir/passes/common/_c_api_utils.py	Avoids asserting on const_value and restores unloaded initializers safely after ONNX API calls.
src/onnx_ir/passes/common/shape_inference_test.py	Adds a test case for shape inference with an initializer that has no tensor data loaded.