Fix split_predict generating invalid ONNX models with missing elem_type

[Copilot]

The split_predict pass generates invalid ONNX models when intermediate values lack explicit value_info. Graph inputs end up with UNDEFINED elem_type (value 0), causing validation failures in onnx.checker and ONNX Runtime load errors.

Changes

Core fix in onnxoptimizer/passes/split.h:

• Added inferElemType() helper that infers type from producing node's inputs when elem_type is UNDEFINED
• Modified split_predict to use type inference instead of blindly copying UNDEFINED elem_type via copyMetadata()
• Modified split_init to ensure output values have valid elem_type before registration

Type inference heuristic:

// For operators where output type matches input type (Add, Sub, Mul, etc.) // Check producing node's inputs for a known elem_type for (const Value* input : producer->inputs()) { if (input->elemType() != TensorProto_DataType_UNDEFINED) { return input->elemType(); } }

Tests:

• Added Python test: test_split_predict_preserves_elem_type()
• Added C++ test: SplitPredictPreservesElemType

Both tests verify optimized models pass ONNX validation and have valid elem_type for all inputs.

Limitations

Type inference works for common operators where output type matches input type. Operators with different output types (Shape, Cast) require proper value_info in the original model.

Original prompt

---

This section details on the original issue you should resolve

<issue_title>[BUG] The pass "split_predict" generates an invalid optimized ONNX model</issue_title>
<issue_description>split_predict pass generates an invalid ONNX model (missing tensor elem_type) — fails onnx.checker and cannot be loaded by ONNX Runtime

Describe the bug
Hi ONNX Optimizer maintainers, thanks for the project!

When optimizing a valid ONNX model named model.onnx using only the split_predict pass, the resulting model.opt.onnx becomes invalid ONNX. It fails onnx.checker.check_model() with:

onnx.onnx_cpp2py_export.checker.ValidationError: Field 'elem_type' of 'type' is required but missing.

and ONNX Runtime fails to load it:

onnxruntime.capi.onnxruntime_pybind11_state.Fail: ... failed: Invalid tensor data type 0.

This suggests the pass may be producing incomplete/incorrect type information (elem_type) for some tensor(s) in the optimized graph.

Environment

OS: Ubuntu 20.04 LTS Python: 3.9.6 onnx: 1.19.0 onnxruntime: 1.19.2 onnxoptimizer: 0.3.13 

To Reproduce

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

split_predict_repro.tar.gz

tar -xzvf split_predict_repro.tar.gz cd split_predict_repro 

2. Create a Python environment (Python 3.9.6) and install dependencies:

python3.9 -m venv .venv source .venv/bin/activate pip install -U pip pip install -r requirements.txt 

3. Run the split_predict pass on the case model and save model.opt.onnx:

python optimize_model.py --case ./case_00051_seed20654888

This script loads model.onnx and runs optimizer.optimize(model, ["split_predict"]), then saves model.opt.onnx.

3. Validate the optimized model with ONNX checker:

python ./check_model.py

check_model.py loads ./case_00051_seed20654888/model.opt.onnx and runs checker.check_model(m).

4. Try to run the optimized model with ONNX Runtime using stored oracle inputs:

python ./run_optimized_model.py

This script creates an InferenceSession from ./case_00051_seed20654888/model.opt.onnx and feeds inputs from oracle.pkl.

Expected behavior

• model.opt.onnx should remain a valid ONNX model after split_predict optimization.
• onnx.checker.check_model() should pass.
• ONNX Runtime should be able to create an InferenceSession for model.opt.onnx.

Actual behavior

ONNX checker fails:

onnx.onnx_cpp2py_export.checker.ValidationError: Field 'elem_type' of 'type' is required but missing.

ONNX Runtime fails to load:

onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Load model ... failed: Invalid tensor data type 0.

Attachments

I’m attaching:

• requirements.txt (dependency versions)
• optimize_model.py (runs split_predict and saves model.opt.onnx)
• check_model.py (runs ONNX checker on model.opt.onnx)
• run_optimized_model.py (loads model.opt.onnx via ORT and runs with oracle.pkl)
• the minimal failing folder ./case_00051_seed20654888 containing model.onnx, oracle.pkl, and the generated model.opt.onnx.

Thanks for your help!</issue_description>

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

@andife I wonder could you try the newest release v0.3.19, on Mar 3, 2024 or even the current github version?

• Fixes [BUG] The pass "split_predict" generates an invalid optimized ONNX model #221

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.

[sonarqubecloud]

Quality Gate passed

Issues
5 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud