I'm working on a model that requires computing what's currently implemented in torch.bmm, with the difference that the first argument is a batch of sparse matrices rather than dense matrices. Since torch.bmm has nice autodiff and gpu support, it's the closest operation I found in pytorch that fits my current needs and I'm converting my sparse tensors to dense tensors to use it, but it's not optimal and it'd be very useful to have a sparse x dense bmm.

cc @ezyang @gchanan @zou3519 @jerryzh168 @vincentqb