About memory leak

[TLB-MISS]

First of all, thank you for providing good code.

I think there is a memory leak in get_processed_features function in the demo.
Each call consumes memory(about 3GB) far exceeding the return value(about 10MB).
I'm guessing something keeps eating VRAM every time process_features_and_mask is called(Not sure), any ideas on how to fix this?

Thanks.