By default cuda kernels are run asynchronously (you need to call torch.cuda.synchronize()) to block until all launched kernels are done.
By default cuda kernels are run asynchronously (you need to call torch.cuda.synchronize()) to block until all launched kernels are done.