Proposal: burn-sparse crate for pruning, sparse formats, and sparse kernels

[huy209vn]

Hey, I’ve been diving into pruning and sparse inference techniques (like Wanda, SparseGPT, DS-NOT, N:M, block sparsity), and I noticed Burn doesn’t yet have a dedicated place for these tools.

What do you think about adding a burn-sparse crate to cover things like pruning methods, sparse weight formats (CSR/BSR/N:M), sparse tensors, and potentially sparse kernels down the line?

Nothing is fully scoped out yet — just thought I’d throw the idea out there to start the discussion and see if it fits with Burn's direction.

related issues:
Support Sparse Tensors #846

[khoek]

Just found out about the absence and was disappointed myself!

I refer you to: #1998

[laggui]

There was definitely some interest, though since the original issue and draft PR I haven't heard much.

As mentioned in the original discussion, it likely needs to be an extension as we cannot guarantee or force the implementation for all backends.