Fix SplitToSequence with scalar uneven split producing incorrect equal-split output

[Copilot]

When optimizing SplitToSequence with a scalar split that doesn't evenly divide the axis dimension, the optimizer was emitting a Split node with only num_outputs, which produces an equal split — silently corrupting the model semantics.

Example: Input [1, 8400, 80] with split=5000 on axis=-2 should produce [1, 5000, 80] + [1, 3400, 80], but was producing [1, 4200, 80] + [1, 4200, 80].

Changes

• optimizer/_constant_folding.py — In split_to_sequence, when the scalar split doesn't evenly divide the axis dimension (split_dimension_size % split_size != 0), construct an explicit 1-D constant tensor [split_size, ..., remainder] and pass it as the split input to Split. Even splits retain the existing num_outputs-only path.

# Before (always equal split via num_outputs)
split_values = op.Split(input, axis=axis, num_outputs=num_outputs, ...)

# After (explicit sizes for uneven case)
remainder = split_dimension_size - (num_outputs - 1) * split_size
explicit_split = op.Constant(value_ints=[split_size, ..., remainder], ...)
split_values = op.Split(input, explicit_split, axis=axis, num_outputs=num_outputs, ...)

• optimizer/_constant_folding_test.py — Adds a regression test asserting the generated Split node carries explicit split sizes [5000, 3400] rather than performing an equal split.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>SplitToSequence layer with unequal split generates incorrect onnx model</issue_title>
<issue_description>converting the attached ONNX model

with the following script:

from onnxscript import ir
from onnxscript.optimizer import optimize

ir_model = ir.load("input.onnx")
optimized_model = optimize(ir_model)
ir.save(optimized_model, "output.onnx")

results in the following output model:

The uneven split in the SplitToSequence layer has been converted to an even split Split layer.

input.onnx.zip

versions of relevant packages:

ml_dtypes         0.5.4
onnx              1.20.1
onnx-ir           0.2.0
onnxscript        0.6.2
```</issue_description>

<agent_instructions>Try fix this</agent_instructions>

## Comments on the Issue (you are @copilot in this section)

<comments>
</comments>

• Fixes SplitToSequence layer with unequal split generates incorrect onnx model #2857

---

📱 Kick off Copilot coding agent tasks wherever you are with GitHub Mobile, available on iOS and Android.