Segmentation Fault in `torch.ops.aten.matrix_exp_backward`

[WLFJ]

🐛 Describe the bug

example:

import torch

def f(*args):
    sym_0, sym_1, sym_2, sym_3, sym_4, sym_5, sym_6 = args
    var_976 = torch.ops.aten.blackman_window(window_length= sym_0, periodic= sym_1)
    var_956 = torch.ops.aten.special_logsumexp(self= var_976, dim= sym_2, keepdim= sym_3)
    var_781 = torch.ops.aten.randint(low= sym_4, high= sym_5, size= sym_6)
    print(var_956, var_781)
    return torch.ops.aten.matrix_exp_backward(self= var_956, grad= var_781)

f(358, False, (-1,), False, -1, 0, (1,))

result:

tensor(6.3650) tensor([-1])
[W209 20:23:45.571710835 TensorShape.cpp:4475] Warning: Tensor.mH is deprecated on 0-D tensors. Consider using x.conj(). (function operator())
fish: Job 2, 'python3 sigsegv-matrix_exp_back…' terminated by signal SIGSEGV (Address boundary error)

Versions

pytorch 2.7.0.dev20250209+cu124

cc @malfet

[mansiag05]

I would like to work on the issue.
The segfault is arising from backward_analytic_function_of_a_matrix function, as it is working with scalar as input eventually causing segfault in
self_transposed_sizes[self.dim() - 2] <<= 1; self_transposed_sizes[self.dim() - 1] <<= 1;

Minimum reproducible code:

import torch

A = torch.tensor(0.23565612)
B = torch.rand((1,2))

torch.ops.aten.matrix_exp_backward(A, grad=B)

[WLFJ]

Thanks for taking a look! You can go ahead and start working on this issue. I’m happy to provide any help or additional information you might need along the way.