Hello,
I am setting a new environment from PyTorch 1.6 to PyTorch 1.11.
My major focus is to extend C++ with some ATen functions.
With PyTorch 1.6, I built my custom Conv2d and Linear functions referring to https://github.com/jordan-g/PyTorch-cuDNN-Convolution and https://pytorch.org/tutorials/advanced/cpp_extension.html.
The first reference was very helpful, from which I used the three functions of at::cudnn_convolution, at::cudnn_convolution_backward_input, and at::cudnn_convolution_backward_weight.
With the PyTorch 1.6, training models with my cpp extension operates well.
However, after migrating to PyTorch 1.11, error occured when calling the last two functions for backwards, i.e., at::cudnn_convolution_backward_input and at::cudnn_convolution_backward_weight; calling the first function, at::cudnn_convolution was fine.
The error message is error: 'cudnn_convolution_backward_input' is not a member of 'at'.
I installed both PyTorch 1.6 and 1.11 via anaconda with the same cuda toolkit of 10.2.
I found that the two backward functions are in the library of PyTorch 1.6 by searching the functions with the command of grep -rnw './' -e 'cudnn_convolution_backward_input' @ .../python3.8/site-packages/torch/.
Meanwhile, in the PyTorch 1.11 library, I could not find the two backward functions.
How can I use the functions in the latest version?
Thank you.