[BUG] Pass "eliminate_nop_concat" causes output mismatch

onnxoptimizer eliminate_nop_concat causes output mismatch

Issue
Hi, maintainers. Running the single pass eliminate_nop_concat with onnxoptimizer 0.3.19 changes numerical outputs. Two real models regress after this pass only, while the original models run correctly with the same oracle inputs.

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

eliminate_nop_concat_repro.tar.gz

tar -xzvf eliminate_nop_concat_repro.tar.gz
cd eliminate_nop_concat_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_nop_concat:

• python optimize_model.py --case ./case_00531_seed24456008
• python optimize_model.py --case ./case_03379_seed47009320
  (Writes model.opt.onnx next to model.onnx in each case directory.)

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

• python diff_test.py --case ./case_00531_seed24456008
• python diff_test.py --case ./case_03379_seed47009320

Observed results

• case_00531_seed24456008: overall max_abs=1.953e-03, max_rel=5.995e-04 (one output tensor differs; others match).
• case_03379_seed47009320: large drift, overall max_abs=1.614e+04, max_rel=1.918e+01; multiple outputs diverge.
• ORT logs warnings about missing constant folding kernels (Cos/Floor/Tan/Resize); these are present for both original and optimized runs and are not fatal, but optimized outputs still mismatch.

Expected
eliminate_nop_concat should be semantics-preserving. Applying this pass alone should not change any output values. Please investigate why the optimized graphs for the above cases produce different outputs and whether the pass is incorrectly modifying concat inputs/axes or tensor ordering.

Differential Test Output Details

1. case_00531_seed24456008

Case: case_00531_seed24456008
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 2, 1)
  output[1]: max_abs=0.000e+00, max_rel=0.000e+00, shape=()
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 1, 1)
  output[3]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(7, 23, 56, 1, 2)
  output[4]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 2)
  output[5]: max_abs=0.000e+00, max_rel=0.000e+00, shape=()
  output[6]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(7, 23, 56, 1, 1)
  output[7]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(7, 23, 56, 1, 1)
  output[8]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(7, 23, 1, 1, 1)
  output[9]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(9016,)
  output[10]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(7, 23, 56, 1, 1)
  output[11]: max_abs=1.953e-03, max_rel=5.995e-04, shape=(7, 23, 56, 1, 1)
Overall: max_abs=1.953e-03, max_rel=5.995e-04

2. case_03379_seed47009320

Case: case_03379_seed47009320
  output[0]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1)
  output[1]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(32, 1)
  output[2]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 42, 41, 1, 32)
  output[3]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 41, 1, 32)
  output[4]: max_abs=0.000e+00, max_rel=0.000e+00, shape=(1, 1, 41, 1, 32)
  output[5]: max_abs=1.191e-01, max_rel=1.284e+00, shape=(1, 45, 41, 1, 32)
  output[6]: max_abs=1.614e+04, max_rel=1.918e+01, shape=(1, 45, 41, 1, 32)
  output[7]: max_abs=1.225e+01, max_rel=1.284e+00, shape=(45, 41, 1, 32)
  output[8]: max_abs=8.339e-01, max_rel=1.322e+01, shape=(45, 41, 1)
  output[9]: max_abs=7.356e-01, max_rel=1.224e+01, shape=(45, 41, 1)
  output[10]: max_abs=8.339e-01, max_rel=1.322e+01, shape=(45, 41, 1)
  output[11]: max_abs=8.339e-01, max_rel=1.322e+01, shape=(45, 41, 1, 1)
Overall: max_abs=1.614e+04, max_rel=1.918e+01

Attachments

• README.md (this document)
• requirements.txt (dependency versions)
• optimize_model.py (runs only eliminate_nop_concat and saves model.opt.onnx)
• diff_test.py (runs original vs optimized with oracle inputs and reports max_abs/max_rel)
• case_00531_seed24456008/ (contains model.onnx which is original model, oracle.pkl which is the input data of both original and optimized model)
• case_03379_seed47009320/ (contains model.onnx, oracle.pkl)

Start by running optimize_model.py and diff_test.py for both listed case directories with the pinned dependencies, then compare each original model.onnx with its generated model.opt.onnx. Trace how eliminate_nop_concat changes the affected graphs and verify that both cases remain numerically equivalent to the original outputs when the pass is applied alone.