It seems that x.index_fill_() can change memory outside x, when x is a cuda tensor.
If x is non-cuda tensor, we get:
>>> import torch
>>> x = torch.Tensor([1,1,1])
>>> x.index_fill_(0, torch.LongTensor([100]), -1)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
RuntimeError: invalid argument 2: out of range at /opt/conda/conda-bld/pytorch_1503966894950/work/torch/lib/TH/generic/THTensor.c:861
In contrast, when x is cuda tensor, index_fill_() does not make any error
>>> a = torch.Tensor([1,1,1]).cuda()
>>> a.index_fill_(0, torch.LongTensor([100]).cuda(), -1)
1
1
1
[torch.cuda.FloatTensor of size 3 (GPU 0)]
It's hard to share the whole code, but I have noticed that such operation outside a tensor did affect the performance of existing network, so I'm afraid that this op can change arbitrary memory on GPU which can be dangerous. Could you check this out?
It seems that
x.index_fill_()can change memory outside x, when x is a cuda tensor.If x is non-cuda tensor, we get:
In contrast, when x is cuda tensor,
index_fill_()does not make any errorIt's hard to share the whole code, but I have noticed that such operation outside a tensor did affect the performance of existing network, so I'm afraid that this op can change arbitrary memory on GPU which can be dangerous. Could you check this out?