Bowen Bao

> @BowenBao : for your awareness: this PR proposes to promote the axes attribute to input for various Reduce* ops. (This was already done for ReduceSum previously, but not for...

The issue is more suited to be filed under onnxruntime repo. Looks like an additional check for NaN is needed on line https://github.com/microsoft/onnxruntime/blob/33578cc76efc19b50c9fc011215b2777de193cd1/onnxruntime/core/providers/cuda/math/clip_impl.cu#L14 cc @yuslepukhin to check if my understanding...

> Can this operator be constructed using existing onnx operators? Composition of `Where`, `Greater`, `Floor`, `Ceil` perhaps?

We are developing a validation tool `find_mismatch` to help investigate this kind of issues. Would you mind trying https://github.com/pytorch/pytorch/blob/ef495b7d6453da445861a105b3cc6f24d840da3a/torch/onnx/verification.py#L1723 via pytorch master/nightly build? The doc can be found at https://pytorch.org/docs/master/onnx.html#torch.onnx.verification.find_mismatch.

Hi @veelken, thanks for the detailed response. Could you try `find_mismatch(torch_model, (data, ))`? The `input_args` argument of `find_mismatch` needs to be a tuple of tensors.

@veelken , glad you found the root cause! It is a known issue #32960 that `scatter_add` may be unstable for opset version < 16. Would you mind give a try...

It looks like the failure is related after all, and it is similar/same to the failure exhibited in #91059. Both PRs adjusted frame skipping behavior for break in loops. But...

@pytorchbot rebase

@ezyang so far we only know test/test_torch.py test_scalar_check needs skipping. I updated #91059 (seems no write access to this PR) with the skip and will add more skips if discovered.

@ezyang final list of tests skipped https://github.com/pytorch/pytorch/pull/91059/files#diff-8aa1a200ec63d23db422aa31b6dca1e6cb372887c43b064ef435210b1b0dec0a If it looks good I'll go ahead and merge.