[Bug]: Batched Matrix-Vector Multiplication does not work correctly

[FOsterfeld]

What happened?

During testing, I found that PyTorch's matmul does not support batched matrix-vector multiplication. If an argument has more than two dimensions, the shapes, including the batch dimensions, are simply broadcasted and it is treated as a batched matrix-matrix multiplication. As a result, batched matrix-vector multiplication (#1261) does not work in all cases.

Since this is not supported by PyTorch itself, and since one can just use unsqueeze/expand_dims to handle this, I would suggest removing this functionality from Heat, and maybe raising a NotImplementedError.

Relevant link: Matrix-vector multiply (handling batched data) - PyTorch Forums

Code snippet triggering the error

No response

Error message or erroneous outcome

No response

Version

main (development branch)

Python version

None

PyTorch version

None

MPI version

No response

[github-actions]

Branch bugs/1645-_Bug_Batched_Matrix-Vector_Multiplication_does_not_work_correctly created!