[Training] Convert to onnx with opset 16 but cannot find an implementation for operator Cos with opset 7.

[SunHaoOne]

Describe the issue

I have convert the onnx with opset 16, but cannot run it with ort. It said onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for Cos(7) node with name '/encoder/Cos'
I have provided more details in this issue. Could you please give me some ideas?

To reproduce

#19414

Urgency

No response

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.16.3

PyTorch Version

2.1.1+cu118

Execution Provider

Default CPU

Execution Provider Library Version

cu118

[baijumeswani]

Are you able to share the onnx model?
Or if you're able to share the all the steps such that I can reproduce this on my end, I can take a deeper look.

[SunHaoOne]

Are you able to share the onnx model? Or if you're able to share the all the steps such that I can reproduce this on my end, I can take a deeper look.

Hi, @baijumeswani .Thanks for your reply, I update the onnx file here. By the way, the model can be successfully run by tensorRT.

[baijumeswani]

I think the problem seems to be that the model is using Cos for inputs with types float64:

The problem is that onnxruntime has not defined a cpu kernel for Cos that takes in tensors of type double. See here.

Are you able to change your model to use float 32 inputs?

[SunHaoOne]

The problem is that onnxruntime has not defined a cpu kernel for Cos that takes in tensors of type double. See here.

Are you able to change your model to use float 32 inputs?

Hi, @baijumeswani.
Thank you very much for your insightful suggestions. I have checked the inputs; they are float64, and if we change it to float32, it does not matter for my problems. So, when I exported the model to ONNX, I only needed to change the dummy_input type from float64 to float32, and it works. Thank you again!