Indexing.cu error when use_deterministic_algorithms is True

1

To reproduce the bug:

import torch
torch.use_deterministic_algorithms(True)
a = torch.full(torch.zeros(2,10).shape, 0.1, device='cuda:0')
b = (torch.tensor([0, 1, 1]), torch.tensor([1,2,3]))
a[b] = 1
print(a)

image
image2160×148 67.7 KB

Without the second line the code would exit normally.

2

Could you update to the latest nightly release, as this issue should be solved already?

1 Like