@ThomasDelteil `cudnn_off` is more of a testing functionality which allows us to compare all versions (in-house CPU, in-house GPU, CuDNN) without re-building the package, but sometimes if user is not satisfied with CuDNN for some reason they can also use that in their code. I think you got the idea behind the change to GPU impl. I'm not familiar with the operator either but we know we should give same results from all 3 version and the GPU one is violating that rule, so after comparison I found that GPU code is a bit different, thus the change.
[ Full content available at: https://github.com/apache/incubator-mxnet/pull/12557 ] This message was relayed via gitbox.apache.org for [email protected]
