[BUG] Pass “split_init“ produces invalid model (missing value_info.type)

Pass “split_init“ produces invalid model (missing value_info.type)

Issue
Running the single pass split_init with onnxoptimizer 0.3.19 makes the optimized model fail ONNX validation: ValidationError("Field 'type' of 'value_info' is required but missing."). The optimizer falls back to the original model. Differential testing shows outputs remain identical, indicating the failure is structural (missing type info) rather than numerical drift.

Environment

• Ubuntu 20.04
• Python 3.10
• onnx==1.19.0
• onnxruntime==1.23.2
• onnxoptimizer==0.3.19 (latest)

Repro steps (run from this folder)

1. Download and unzip the attached archive, then cd into the extracted directory

split_init_repro.tar.gz

tar -xzvf split_init_repro.tar.gz
cd split_init_repro

2. Create a Python environment (Python 3.10) and install dependencies:

python3 -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

3. Optimize the case with only split_init:

• python optimize_model.py --case ./case_00114_seed21153785
  • Warning observed during optimize: optimized model failed validation (ValidationError("Field 'type' of 'value_info' is required but missing.")); falling back to original model
  • The optimized model is invalid due to missing value_info.type.

4. Differential test original vs optimized outputs using stored oracle inputs:

• python diff_test.py --case ./case_00114_seed21153785

Observed results

• python optimize_model.py --case case_00114_seed21153785 prints: Warning: optimized model failed validation (ValidationError("Field 'type' of 'value_info' is required but missing.")); falling back to original model
• Differential test:
  • Case: case_00114_seed21153785
  • Each output node is identical (max_abs=0, max_rel=0)

Expected
split_init should produce a valid, checkable model. The pass currently emits a graph missing value_info.type, causing onnx.checker to fail even though numerical outputs match the original. Please investigate how the pass handles value_info when splitting the graph.

Differential Test Output Details

Case: case_00114_seed21153785
All outputs are identical (max_abs=0, max_rel=0)

Attachments

• README.md (this document)
• requirements.txt (dependency versions)
• optimize_model.py (runs only split_init and saves model.opt.onnx)
• diff_test.py (runs original vs optimized with oracle inputs; prints equality when all outputs match)
• case_00114_seed21153785/ (contains model.onnx and oracle.pkl used for both runs)

Reproduce the failure with optimize_model.py using the split_init pass, then inspect the optimizer code handling value_info during graph splitting. Use diff_test.py to confirm outputs remain identical, and verify completion with onnx.checker accepting the optimized model without a missing value_info.type error.