Handling initializers in GraphBuilder

[gramalingam]

Motivation: Use of some standard graph-building techniques in Mobius fail when used to create ir.Function, because ONNX Functions do not allow initializers. This PR helps address this. It also reduces some of the boiler-plate code in creating functions via an utility function.

(a) Update the logic for creating initializers to ensure that they are always added to the root (main) Graph.
(b) Add a utility to convert initializers into Constant nodes (which is necessary for a Function in ONNX).
(c) Add a utility to build a function

Note: An alternative considered was adding an option to the GraphBuilder so that we automatically construct Constant nodes instead of initializers when the options says so. While that avoids the extra pass in the end, it has some minor implications to what the graph would look like (in terms of whether we want all Constant nodes upfront in one place, and what kind of node-names (based on node numbering) we generate etc. Hence, leaving it in the current form. But it can be changed as above if desirable.

Usage: With the utility function, we should be able to create ir.Functions as below (example from Mobius).

def causal_conv_nd_with_state(*, kernel_size, channels, ndim, activation):
    def body(op, input_val, weight_val, bias_val, conv_state_val):
        # ... body (unchanged) ...
        return output, present_state

    return builder.build_function(
        body,
        [
            ir.Value(name="input"),
            ir.Value(name="weight"),
            ir.Value(name="bias"),
            ir.Value(name="conv_state"),
        ],
        domain=DOMAIN,
        name="CausalConvWithState",
        attributes=[ir.Attr("activation", ir.AttributeType.STRING, activation)],
        opset_imports={"": OPSET_VERSION},
    )

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates GraphBuilder so constant promotion (and explicit initializer() calls) registers initializers on the root graph for subgraph visibility, and adds a post-pass utility to convert initializers to Constant nodes for ONNX ir.Function bodies.

Changes:

• Add lift_initializers_to_constants(graph) to replace graph initializers with Constant nodes (skipping input-default initializers).
• Move constant caching to the root builder so sibling subgraphs share promoted constants.
• Add unit tests validating root initializer placement and the lifting behavior.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.

File	Description
onnxscript/_internal/builder.py	Adds initializer→Constant lifting utility; routes constant creation/initializer registration through the root builder for subgraph visibility.
onnxscript/_internal/builder_test.py	Adds tests covering root initializer storage, cache sharing across subgraphs, and lifting initializers into Constant nodes for function bodies.

[codecov]

Codecov Report

❌ Patch coverage is 96.55172% with 10 lines in your changes missing coverage. Please review.
✅ Project coverage is 72.48%. Comparing base (63ffecf) to head (09098be).
⚠️ Report is 1 commits behind head on main.
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
onnxscript/_internal/builder.py	89.79%	5 Missing and 5 partials ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2889      +/-   ##
==========================================
+ Coverage   72.30%   72.48%   +0.17%     
==========================================
  Files         241      241              
  Lines       29678    29915     +237     
  Branches     2916     2935      +19     
==========================================
+ Hits        21459    21684     +225     
- Misses       7227     7233       +6     
- Partials      992      998       +6     

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

[Copilot]

Pull request overview

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.