Open riccardopinosio opened 1 week ago
This looks similar to the issue I had and fixed in https://github.com/microsoft/onnxruntime/pull/22414 . You can verify it's the same issue if you change your loss to crossentropy and see artifact generation succeed. If so, if you try using a nightly build, local build using master, or wait for the 1.20 release, it should be resolved (with mse loss).
Describe the issue
Hello,
see also this discussion. I'm opening this one as I think it's an issue as sifting through previous issues training should work for bert models.
I am trying to generate artifacts for distilbert like so:
The exported onnx model works perfectly for inference, but artifact generation throws up:
Seems to have issues building the gradient graph as it gets out of bounds on OutputDefs.
To reproduce
See the code provided above.
Urgency
It's blocking the development of go bindings to onnx training which we want to use in our product.
ONNX Runtime Installation
Released Package
ONNX Runtime Version or Commit ID
1.19.2
PyTorch Version
2.4.1+cu121
Execution Provider
Default CPU
Execution Provider Library Version
No response