The examples go out of memory. Probably we need some `CUDA.reclaim()`?
The examples go out of memory. Probably we need some
CUDA.reclaim()?