Disabling cudnn algorithms

Hi, is there a PyTorch equivalent for the following and alike?

https://docs.nvidia.com/deeplearning/frameworks/tensorflow-user-guide/index.html#tf_enable_winograd_nonfused

I wonder, in TF it is integrated into the framework,
but maybe it is possible to do the same from the cudnn dynamic lib?

No, currently you cannot disable specific cudnn algorithms manually. We are working on the enablement of the cudnn v8 API, which allows more flexible filtering.

1 Like