torch.sspaddmm should broadcast the input tensor

[ilovepytorch]

🐛 Bug

torch.sspaddmm(input, mat1, mat2) is equivalent with torch.addmm(input, mat1, mat2) and torch.sparse.addmm(input, mat1, mat2) except input and mat1 are sparse. So it should broadcast the input tensor like the other two.

To Reproduce

import torch
input = torch.rand([1, 1])
mat1 = torch.rand([2, 3])
mat2 = torch.rand([3, 3])
res1 = torch.addmm(input, mat1, mat2)
print("addmm pass")
input = input.to_sparse()
mat1 = mat1.to_sparse()
res2 = torch.sspaddmm(input, mat1, mat2)

Result:

addmm pass
Traceback (most recent call last):
    res2 = torch.sspaddmm(input, mat1, mat2)
RuntimeError: sspaddmm: Argument #1: Expected dim 0 size 2, got 1

Expected behavior

It should broadcast the input tensor and not raise this kind of error.

Environment

• PyTorch Version (e.g., 1.0): 1.10.0, 1.9.0
• OS (e.g., Linux): Linux
• How you installed PyTorch (conda, pip, source): conda
• Python version: 3.9.5
• CUDA/cuDNN version: 11.1
• GPU models and configuration: RTX 3090

cc @nikitaved @pearu @cpuhrsch @IvanYashchuk

[ilovepytorch]

Is there any update?

[IvanYashchuk]

Broadcasting is not implemented for sparse tensors because the tensor in the sparse case needs to be materialized and repeat all the values making it much larger. I guess there's no operation that currently supports sparse tensors and implements broadcasting behavior.
In the future, broadcasting behavior could be implemented for sparse tensors, but I don't think that anybody is actively working on this at the moment.

[cpuhrsch]

@IvanYashchuk - Pearu did some work for sparse specific behavior for broadcast_to, but the underlying operation really is expand.

@ilovepytorch - Given all the other operations that are still missing for Tensors with sparse Layouts this might show up a little later.

[ilovepytorch]

Thanks for your answers!