Moving a tensor to cuda device cause illegal memory access in Pytorch

user_8244472 0
Viewed 5778

I am trying the following snippet in Colab but causes the following error. Is it wrong to move a tensor object to Cuda device?.

import torch
a = torch.Tensor(torch.randn(5,5,5))
# a.device("cuda")
device = torch.device("cuda")
class abc(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.w1 = torch.nn.Linear(5,5)

    def forward(self,x):
        return self.w1(x)
mod = abc()
a.cuda()
mod.to(device)
mod(a)

Output:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-10-5372fb4d5512> in <module>()
     11         return self.w1(x)
     12 mod = abc()
---> 13 a.cuda()
     14 mod.to(device)
     15 mod(a)

RuntimeError: CUDA error: an illegal memory access was encountered
pytorch google-colaboratory deep-learning python
2 Answers

This works for me on Google colab:

import torch
a = torch.randn(5,5,5)
a = a.to("cuda") # or just a = torch.randn((5,5,5), device='cuda')

class abc(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.w1 = torch.nn.Linear(5,5)

    def forward(self,x):
        return self.w1(x)
mod = abc()
mod.to("cuda")
mod(a)

Output:

tensor([[[ 1.5691e+00,  8.0326e-01,  1.4352e+00,  7.3295e-01,  3.2156e-01],
         [ 5.1630e-01, -2.2816e-03,  7.1052e-01,  1.9250e-01,  8.3110e-01],
         [ 7.6572e-01, -8.9701e-01,  2.7974e-01,  7.4309e-04,  9.5218e-01],
         [ 2.0723e-01, -1.0049e+00,  1.6938e+00,  1.0019e+00,  7.9305e-01],
         [-1.0973e-02, -1.1260e-01,  1.0521e+00, -1.3839e-01, -4.2380e-01]],

        [[ 1.3870e+00,  1.1620e+00, -3.6523e-01, -5.6704e-01,  4.2481e-01],
         [ 1.6204e-01,  8.3231e-02, -5.9607e-01, -1.0912e+00, -6.1651e-01],
         [ 2.3584e-01, -5.9825e-01,  1.1670e+00,  9.3185e-01,  4.0269e-01],
         [ 1.3120e+00,  1.3967e-01, -5.5048e-01, -9.8143e-01,  3.5059e-01],
         [ 8.0019e-01, -1.8983e-02,  2.3792e-01, -5.9157e-01,  3.5816e-01]],

        [[ 3.9709e-01, -8.7349e-01, -2.9742e-01, -3.8732e-01, -1.7191e-03],
         [-8.7056e-01, -8.8214e-01,  1.0647e+00,  7.7785e-01,  6.3816e-01],
         [ 7.4920e-01, -4.0143e-01,  5.9780e-01,  2.7842e-01,  8.1991e-01],
         [-5.9389e-02, -4.9465e-01, -3.7322e-03, -7.0475e-01, -2.5955e-01],
         [ 1.5722e+00,  6.4410e-01, -5.1310e-02, -1.2716e+00, -1.4607e-01]],

        [[ 6.5152e-02, -6.8772e-01,  1.0366e+00, -2.4278e-01, -2.7106e-01],
         [ 7.0832e-01,  1.4581e-01,  1.9924e-01, -4.1930e-01,  4.0567e-01],
         [ 3.9120e-01, -1.0099e+00,  1.6907e+00,  7.2674e-01,  6.5285e-01],
         [-1.3191e-01, -8.6324e-01, -1.2734e-01, -5.6135e-01, -4.1949e-01],
         [ 5.4183e-02, -5.6837e-01,  5.1347e-02, -5.3199e-01,  2.2167e-01]],

        [[ 9.9237e-02, -5.8725e-01, -3.3042e-01, -8.7371e-01, -2.3261e-01],
         [ 5.5485e-01, -3.5022e-01,  1.1516e-01,  3.8139e-02,  4.6032e-01],
         [-7.5111e-01, -9.7203e-01,  1.7809e-01,  2.2506e-01,  3.6540e-02],
         [ 2.5590e-01,  3.0592e-01,  6.8972e-01,  1.8452e-01,  6.7794e-01],
         [-7.6091e-01, -1.3956e+00,  7.8801e-01, -1.7489e-01, -1.0143e+00]]],
       device='cuda:0', grad_fn=<AddBackward0>)
user_58849550

actually if we run the code, we may get the result if we run the code here.

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

and then, if we enter a.cuda(), we will get

RuntimeError: CUDA error: an illegal memory access was encountered

So may the tensor forward in the model, but it failed, it could be handled in some computation, and then it's not able to transfer to cuda again.

I guess so, but I do not know what the model(nn.Module) forward do to the tensor.

BTW, this is a simple problem here, but if we met in some complex problem, we can find if we have make the same mistake here (forward tensor to cuda again)

user_159584140
Related