How to set MXNET_CUDNN_AUTOTUNE_DEFAULT=0 in Sagemaker

[MikkelWorkF]

Hi

We are using AWS Sagemaker to run our training, and are moving taking the step to GPU.
This messages comes up sometimes
Running performance tests to find the best convolution algorithm, this can take a while... (set the environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
And it stalls for hours. How do I disable this in Sagemaker? It is not possible to just set environment variables.

Thank you.