than adding them.  This way, the gradient buffer can be used as scratch space
during the loss computation.

runs on the GPU, and made affine_transform() take only tensors.

use either CPU or GPU.  Fixed a bug in gemm().

form.

in-place.