How to construct sparse tensor on gpu efficiently?

Hello. Suppose I have a weight tensor with shape n * c * h * w * 11 * 11, which means for every pixel I have 121 weight. Then I want to use those weight to manipulate a image like convolution but every pixel has different kernel weight. Maybe a straightforward solution is to construct a sparse tensor using the weight tensor and them achieve the convolution by matrix multiplication. Therefore, I need to unpack the 121 weight for every pixel to the sparse tensor efficient. Is there a built-in solution in pytorch ?