Assign torch.cuda.FloatTensor to List tensor

[IAMAl]

❓ Questions and Help

I make custom cell and use eager execution on CUDA.
Then I meet segmentation fault to assign torach.cuda.FloatTensor to List tensor.

List Tensor is defined in init(self) as;

self.x     = [torch.zeros(NUM_INPUT, device='cuda') for _ in range(NUM_HIDDEN)]

And the self.x is done as;

for x, t in dataloader_train:
  for time in range(TIME_STEPS):
     for index in range(NUM_HIDDEN-1, 0, -1):
       model.x[index] = model.x[index - 1]
  
     model.x[0] = torch.cuda.FloatTensor(x[0][0][time])

Then output is

Segmentation fault

How to avoid from the fault?

[ptrblck]

Additional information:
Issue was reported in the discussion board and the Segmentation fault should be avoided using model.x[0] = x[0][0][time].cuda() as described here.

That being said, this particular error is created by simply calling:

import torch

torch.cuda.FloatTensor(torch.randn(1))

Using gdb to get the stack trace:

(gdb) run
Starting program: /home/ptrblck/anaconda3/envs/pytorch_latest/bin/python cuda_tensor_creation.py
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".

Program received signal SIGSEGV, Segmentation fault.
__strlen_avx2 () at ../sysdeps/x86_64/multiarch/strlen-avx2.S:62
62	../sysdeps/x86_64/multiarch/strlen-avx2.S: No such file or directory.

Tested using PyTorch 1.3.0-dev20190827 with cudatoolkit=10.0.

[IAMAl]

Again, thank you very much, ptrblk-san!
This experience is my valuable one.

[kcy1019]

@ptrblck Thanks! I've just encountered the same problem and your .cuda() tip worked like a charm :)