Support for Sparse Neural Networks in ONNX-MLIR

[shreya-um]

I’m interested in exploring support for sparse neural networks in ONNX-MLIR — particularly in improving the representation of sparsity in weight matrices.

At the moment, it seems that weight arguments are treated the same way for both dense and sparse matrices:

• Is there any ongoing work to enhance sparsity representation in ONNX-MLIR?
• If not, I’d be interested in contributing to this effort.
• Are there any references, ongoing discussions, or design documents that could help me get started?