I get the following error:
RuntimeError: handle_0 INTERNAL ASSERT FAILED at "../c10/cuda/driver_api.cpp":15, please report a bug to PyTorch.
I get this error when I try to run a transformer model I implemented following this tutorial: An even more annotated Transformer :: pi-tau
The error only occurs when I call LogSoftmax or Softmax over the final outputs, not if I skip softmax altogether.
Package versions:
Torch==2.1.0
Cuda==12.1
GPU = RTX 3060ti
When I lower the dimensionality of the model the error does not appear. I suppose this has got to do with that my GPU runs out of memory then?
Does this mean that Softmax and LogSoftmax are very computationally expensive as they seem to be what finally causes the error?
Ps. This is my first post here, let me know if I shold change/add anything. Thanks!