Issue with random negative numbers on GPU

1

There’s a reported issue with random_ when the lower bound is negative for cuda.LongTensors. However those large numbers seem to be valid indices.

>>> x1 = torch.linspace(1, 10, 10).cuda();

>>> x2 = torch.linspace(21, 30, 10).cuda();

>>> torch.stack((x1, x2))

    1     2     3     4     5     6     7     8     9    10
   21    22    23    24    25    26    27    28    29    30
[torch.cuda.FloatTensor of size 2x10 (GPU 0)]

>>> idx = torch.cuda.LongTensor(3).random_(-10, 0).add_(10); idx

 4.2950e+09
 4.2950e+09
 4.2950e+09
[torch.cuda.LongTensor of size 3 (GPU 0)]

>>> torch.stack([x1[idx], x2[idx]])

  3   5   2
 23  25  22
[torch.cuda.FloatTensor of size 2x3 (GPU 0)]

Is it safe to use those numbers for indexing?

2

Thanks for reporting! This is discussed and being fixed here: https://github.com/pytorch/pytorch/issues/6338

1 Like