RuntimeError: PyTorch convert function for op 'new_empty' not implemented.

[RahulBhalley]

🐞Describing the bug

Hey there,

While converting a PyTorch model, we're getting the error written in title. We don't use torch.new_empty anywhere in the forward method of model but traced_model.graph shows the usage of new_empty op of PyTorch.

To Reproduce

x = torch.rand(1, 3, 512, 512)
traced_model = torch.jit.trace(model, x)
inputs = [ct.ImageType(name='image', shape=x.shape)]
mlmodel = ct.convert(traced_model, inputs=inputs)

Throws the following error:

RuntimeError: PyTorch convert function for op 'new_empty' not implemented.

How can we fix it???

System environment (please complete the following information):

• coremltools version: 6.0b1
• OS (e.g. MacOS version or Linux type): macOS 15.5
• Any other relevant version information (e.g. PyTorch or TensorFlow version): PyTorch 1.12

Additional context

N/A

[aseemw]

That means some part of the pytorch code is getting lowered into the new_empty op after tracing. Can you please check by printing traced_model.forward.graph which part of the code corresponds to the new_empty op?

[TobyRoseman]

Currently we handle the empty op by just aliasing it to the zeros op. To solve this issue, we could probably just alias new_empty to new_zeros. We already support new_zeros.

@RahulBhalley - you could try changing this line to:

@register_torch_op(torch_alias=["new_empty"])

Let us know if that works for you.

[RahulBhalley]

Thanks for quick replies @aseemw and @TobyRoseman. And yes @TobyRoseman, your solution works for new_empty op. But I've encountered issue with non-implementation of torch.normal_ op. Will fix it next. Thanks.

[TobyRoseman]

@RahulBhalley - Thanks for letting us know that works. Let's leave this issue open until we have this fix in main.

[cnpcshangbo]

Thanks for replies @aseemw and @TobyRoseman. Your solution works for new_empty op and torch.normal_ op. But I've encountered issue with non-implementation of profiler::_record_function_enter_new. Can we re-open this issue? Thanks.

Exception has occurred: RuntimeError
PyTorch convert function for op 'profiler::_record_function_enter_new' not implemented.
  File "/Users/boshang/Developer/pytorchMobile/convert2CoreML.py", line 143, in <module>
    mlmodel = ct.convert(
RuntimeError: PyTorch convert function for op 'profiler::_record_function_enter_new' not implemented.

[TobyRoseman]

@cnpcshangbo - profiler::_record_function_enter_new is a separate op. So please create a new GitHub issue for supporting that op. In your issue please include steps to reproduce the problem.