testcase failing due to commit 7b23a1f5d87064daded4b89487dc968fd933051e

[amd-vivekag]

The following commit (7b23a1f) has caused the testcase failure:
https://github.com/iree-org/iree-test-suites/tree/main/onnx_ops/onnx/node/generated/test_averagepool_2d_ceil/

This is blocking bump torchmlir in IREE.

Getting error:

EXEC @test_averagepool_2d_ceil
[FAILED] result[0]: element at index 1 (5) does not match the expected (7.5); expected that the view is equal to contents of a view of 1x1x2x2xf32
  expected:
1x1x2x2xf32=[[[6 7.5][12 13.5]]]
  actual:
1x1x2x2xf32=[[[6 5][8 6]]]

After removing this commit, the error is not there.

[ivangarcia44]

@amd-vivekag I can't reproduce this issue. The test passed:

PASSED onnx/node/generated/test_averagepool_2d_ceil/run_module_io_flags.txt::model.mlir::cpu_llvm_sync

I followed the instructions here:
https://github.com/iree-org/iree-test-suites/tree/main/onnx_ops

The printout above came from this command:

pytest -n auto -rA --timeout=30 --durations=20 --config-files=configs/onnx_ops_cpu_llvm_sync.json --report-log=/tmp/onnx_ops_cpu_logs.json

I made sure that the torch-mlir-opt referred to is the one I just sync'ed, built and verified that has my average pooling change.

$ which torch-mlir-opt
/myfolder/torch-mlir/build/bin/torch-mlir-opt

Can you please let me know which steps you used to reproduce this issue?

[ivangarcia44]

Hi @amd-vivekag, I have not heard back from you in a month. Are things working now? If yes, can this issue be closed?

[amd-vivekag]

Hi @ivangarcia44, I'm sorry, I somehow missed the last month message.

I was seeing the failure in CI with the CL mentioned above. I'll have to check if I was able to reproduce it locally (as far as I remember, I was able to reproduce it).
I'll get back to you with more details.

Thanks for following up. Really appreciate it.

[amd-vivekag]

Hi @ivangarcia44 ,

I'm able to reproduce the issue. I followed following steps:

1. git clone git@github.com:iree-org/iree-test-suites.git
2. cd iree-test-suites/onnx_ops
3. created python environment as per the instructions mentioned in onnx_ops/README.md
4. pip install -r requirements.txt
5. pip install -r requirements-iree.txt
6. Ran following command: pytest -v -n auto -rA --timeout=30 --durations=20 --config-files=configs/onnx_ops_cpu_llvm_sync.json --report-log=/tmp/onnx_ops_cpu_logs.json -k test_averagepool_2d_ceil

You should see following message:

Stdout diagnostics:
EXEC @test_averagepool_2d_ceil
[FAILED] result[0]: element at index 1 (5) does not match the expected (7.5); expected that the view is equal to contents of a view of 1x1x2x2xf32
  expected:
1x1x2x2xf32=[[[6 7.5][12 13.5]]]
  actual:
1x1x2x2xf32=[[[6 5][8 6]]]

Please let me know if you need any other information from my side in this regard.

[ivangarcia44]

@amd-vivekag I am able to reproduce the failure. I am taking a look at it.

[ivangarcia44]

This issue is fixed here: #4144

[ivangarcia44]

The fix for this issue was merged into torch-mlir: #4144. Closing this issue.

[ivangarcia44]

Hi @amd-vivekag, the fix for this issue was merged in torch-mlir: #4144. Can you please close this issue when you get a chance? I can't close it.

[amd-vivekag]

@ivangarcia44 I'm still seeing the issue. Since your changes in torch-mlir are not part of iree build yet, so it might take sometime. Can you please let me know once iree build has your changes?

[ivangarcia44]

@ivangarcia44 I'm still seeing the issue. Since your changes in torch-mlir are not part of iree build yet, so it might take sometime. Can you please let me know once iree build has your changes?

@amd-vivekag I will keep monitoring this and let you know when the change is integrated. Thanks,
Ivan

[ivangarcia44]

Note to find out if the issue is gone. Today, June 23rd, 2025, the bug fix has not made it yet to IREE.

git clone https://github.com/iree-org/iree.git
cd iree
git submodule update --init
find . -name "Pooling.cpp" | xargs grep -nH "doesAvgPoolDivisorNeedsClamping"

If there is a hit above then IREE got the #4144 bug fix. If that is the case, then do the reproduction steps from @amd-vivekag written on April 3.