Convolution with filters of different sizes

Hello, I would like to build a layer which convolves an input image with many filters of different size (5x5, 10x10, 15x15…). As far as I understand, this is not possible using nn.functional.Conv2d. Using a for loop over filters turns out to be slower than simply setting the filter size to be the one of the bigger filter in Conv2d and padding the smaller filters. This seem to involve a lot of unnecessary operations, though. Is there any way to parallelize such convolutions on the GPU?