[BUG] The Pass “eliminate_consecutive_idempotent_ops“ causes output mismatch

[C1ri007]

Pass “eliminate_consecutive_idempotent_ops“ causes output mismatch

Issue
Running the single pass eliminate_consecutive_idempotent_ops with onnxoptimizer 0.4.2 changes numerical outputs. One real model regresses after this pass only, while the original model runs correctly with the same oracle inputs.

Environment

• Ubuntu 20.04
• Python 3.10
• onnx 1.19.0
• onnxruntime 1.19.2
• onnxoptimizer 0.4.2 (latest)

Repro steps (run from this folder)

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

eliminate_consecutive_idempotent_ops_repro.tar.gz

tar -xzvf eliminate_consecutive_idempotent_ops_repro.tar.gz
cd eliminate_consecutive_idempotent_ops_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 a case with only eliminate_consecutive_idempotent_ops:

• python optimize_model.py --case ./case_06833_seed74361546
  (Writes model.opt.onnx next to model.onnx in the case directory.)

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

• python diff_test.py --case ./case_06833_seed74361546

Observed results

• case_06833_seed74361546: mismatch with orginal model after optimization

Expected
eliminate_consecutive_idempotent_ops should be semantics-preserving. Applying this pass alone should not change any output values. Please investigate why the optimized graph for the above case produces different outputs.

Differential Test Output Details

Case: case_06833_seed74361546
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 1, 1, 1)
  output[1]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 1, 4)
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 1, 1)
  output[3]: max_abs=6.079e+00, max_rel=7.709e+01, shape=(5, 1, 57)
  output[4]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 5, 1)
  output[5]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 5, 1, 1)
  output[6]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 5, 1, 1)
  output[7]: max_abs=7.297e-01, max_rel=7.129e+00, shape=(5, 1, 1)
  output[8]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 1, 63)
  output[9]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(15, 1, 63)
  output[10]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5,)
  output[11]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 63)
  output[12]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 63)
  output[13]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(5, 63)
Overall: max_abs=6.079e+00, max_rel=7.709e+01

Attachments

• README.md (this document)
• requirements.txt (dependency versions)
• optimize_model.py (runs only eliminate_consecutive_idempotent_ops and saves model.opt.onnx)
• diff_test.py (runs original vs optimized with oracle inputs and reports max_abs/max_rel)
• case_06833_seed74361546/ (contains model.onnx which is original model, oracle.pkl which is the input data for both original and optimized model)