Update SplitToSequence in constant folding

[titaiwangms]

Split input (SymbolicTensor) could have no const_value, but with shape that gives us information of how many outputs an op.Split should return.

[Copilot]

Pull Request Overview

This PR updates the SplitToSequence function in constant folding to handle cases where the Split input has no constant value but has shape information that can be used to determine the number of outputs. It also refactors the constant folding pass to improve its API and reduce verbosity in logging.

Key changes:

• Enhanced split_to_sequence to use shape information when constant values aren't available
• Replaced should_fold callback parameter with always_fold_ops collection for better API design
• Streamlined logging and code structure in the constant folding pass

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

File	Description
onnxscript/optimizer/_constant_folding.py	Updated SplitToSequence logic, refactored FoldConstantsPass API, improved logging
onnxscript/function_libs/torch_lib/ops/core.py	Added trace_only flag to aten_split_with_sizes function

[Copilot]

There's a typo in the docstring: 'neede' should be 'needed'.

Suggested change

	always_fold_ops: Collection of op types that should always be folded.
	For ops from the default opset, only op_type is neede (e.g. "Transpose"),
	For ops from the default opset, only op_type is needed (e.g. "Transpose"),

[Copilot]

There's a typo in the docstring: 'neede' should be 'needed'.

Suggested change

	always_fold_ops: A collection of op types that should always be folded,
	regardless of their input or output sizes. For ops from the default opset,
	only op_type is neede (e.g. "Transpose"), otherwise specify the domain
	only op_type is needed (e.g. "Transpose"), otherwise specify the domain

[Copilot]

The debug message format string expects one placeholder (%s) but receives two arguments: the node and the list of input names. Either add another %s placeholder or combine the arguments.

Suggested change

	if logger.isEnabledFor(logging.DEBUG):
	logger.debug(
	"Skipping constant folding for node %s because it has non-constant inputs",
	"Skipping constant folding for node %s because it has non-constant inputs: %s",

[justinchuby]

This should be reverted

[codecov]

Codecov Report

❌ Patch coverage is 50.00000% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 63.26%. Comparing base (d98e3dd) to head (ecabee2).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
onnxscript/optimizer/_constant_folding.py	48.57%	13 Missing and 5 partials ⚠️

❗ There is a different number of reports uploaded between BASE (d98e3dd) and HEAD (ecabee2). Click for more details.

HEAD has 1 upload less than BASE

Flag	BASE (d98e3dd)	HEAD (ecabee2)
	21	20

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2542      +/-   ##
==========================================
- Coverage   69.99%   63.26%   -6.73%     
==========================================
  Files         216      216              
  Lines       26074    26073       -1     
  Branches     2618     2619       +1     
==========================================
- Hits        18250    16496    -1754     
- Misses       6921     8745    +1824     
+ Partials      903      832      -71     

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

[titaiwangms]

#2544