Pre-load libcublas and libcudnn

Is there any way to pre-load the libcublas and libcudnn libraries?

I’m working with an inference procedure for single samples and wondering if there is any way to do this since loading takes more time than inference itself.


There are numerous startup overheads in TF such as graph deserialization and compilation, GPU initialization, etc. To avoid these, you need to use a persistent process that waits between inference inputs.

Thanks for the tip, @nluehr