Open tianleiwu opened 3 years ago
@yufenglee , @SherlockNoMad , @ashbhandare - change it to "kMSDomain" ?
@yufenglee , @SherlockNoMad , @ashbhandare - change it to "kMSDomain" ?
agree. let's fix this.
I updated symbolic shape inference (in https://github.com/microsoft/onnxruntime/pull/7606) so that we will not call onnx shape inference on LayerNormalization node.
It is not urgent right now. Let's add this to back log.
@tianleiwu - backlog created for backend team, close this for now with your workaround. We can sync offline on the progress. Thanks for reporting it.
Not fixed yet.
is there a workaround?
Describe the bug
Onnxruntime defined some operators (like LayerNormalization, SimplifiedLayerNormalization etc) in onnx domain: https://github.com/microsoft/onnxruntime/blob/8d737f977056444a307f1b7f0bcd402fba62d790/onnxruntime/contrib_ops/cuda/cuda_contrib_kernels.cc#L70-L75
However, they are not standard onnx operators as listed in (https://github.com/onnx/onnx/blob/master/docs/Operators.md).
That causes some issue. For example, symbolic shape inference outputs warning:
Warning: Unsupported operator LayerNormalization. No schema registered for this operator.
Urgency If there are particular important use cases blocked by this or strict project-related timelines, please share more information and dates. If there are no hard deadlines, please specify none.
None
System information
To Reproduce
Expected behavior How about change to MS domain, and add some logic for backward compatible (like rewrite the domain for those operator during loading models).
Screenshots If applicable, add screenshots to help explain your problem.
Additional context Add any other context about the problem here. If the issue is about a particular model, please share the model details as well to facilitate debugging.