When using the CUDAExecutionProvider with the default configuration, some settings will drastically affect the inference speed.
For example, cudnn_conv_algo_search (default=EXHAUSTIVE) will make a variable input network drastically slower compared to HEURISTIC.
Or it is simply that we just don't want to use Cuda...
This change gives users the ability to pass such options.
When using the CUDAExecutionProvider with the default configuration, some settings will drastically affect the inference speed. For example, cudnn_conv_algo_search (default=EXHAUSTIVE) will make a variable input network drastically slower compared to
HEURISTIC
.Or it is simply that we just don't want to use Cuda... This change gives users the ability to pass such options.