ConvTranspose1d extremely slow on GPU (T4), slower than CPU

[swpd]

Thank you for your explanations!

@ptrblck

Yes, ran it with the default setup first and compared it to benchmark mode, which didn’t yield a speedup.

How to setup pyorch with cudnn8? By compiling from source with system cudnn8 library setup?

@ptrblck

It’s a cudnn “data gradient” kernel, which computes the gradient for the input activation.

Gradients are still computed even with torch.no_grad()?