Differences between `torch.Tensor` and `torch.cuda.Tensor` - pytorch

Question

We can allocate a tensor on GPU using torch.Tensor([1., 2.], device='cuda'). Are there any differences using that way rather than torch.cuda.Tensor([1., 2.]), except we can pass in a specific CUDA device to the former one?
Or in other words, in which scenario is torch.cuda.Tensor() necessary?

Answer 1

So generally both torch.Tensor and torch.cuda.Tensor are equivalent. You can do everything you like with them both.
The key difference is just that torch.Tensor occupies CPU memory while torch.cuda.Tensor occupies GPU memory. Of course operations on a CPU Tensor are computed with CPU while operations for the GPU / CUDA Tensor are computed on GPU.
The reason you need these two tensor types is that the underlying hardware interface is completely different. Apart from the point it doesn't make sense computationally, you will get an error as soon as you try to do computations between torch.Tensor and torch.cuda.Tensor:
import torch
# device will be 'cuda' if a GPU is available
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# creating a CPU tensor
cpu_tensor = torch.rand(10)
# moving same tensor to GPU
gpu_tensor = cpu_tensor.to(device)
print(cpu_tensor, cpu_tensor.dtype, type(cpu_tensor), cpu_tensor.type())
print(gpu_tensor, gpu_tensor.dtype, type(gpu_tensor), gpu_tensor.type())
print(cpu_tensor*gpu_tensor)
Output:
tensor([0.8571, 0.9171, 0.6626, 0.8086, 0.6440, 0.3682, 0.9920, 0.4298, 0.0172,
0.1619]) torch.float32 <class 'torch.Tensor'> torch.FloatTensor
tensor([0.8571, 0.9171, 0.6626, 0.8086, 0.6440, 0.3682, 0.9920, 0.4298, 0.0172,
0.1619], device='cuda:0') torch.float32 <class 'torch.Tensor'> torch.cuda.FloatTensor
---------------------------------------------------------------------------
RuntimeError Traceback (most recent call last)
<ipython-input-15-ac794171c178> in <module>()
12 print(gpu_tensor, gpu_tensor.dtype, type(gpu_tensor), gpu_tensor.type())
13
---> 14 print(cpu_tensor*gpu_tensor)
RuntimeError: Expected object of type torch.FloatTensor but found type torch.cuda.FloatTensor for argument #2 'other'
As the underlying hardware interface is completely different, CPU Tensors are just compatible with CPU Tensor and verse visa GPU Tensors are just compatible to GPU Tensors.
Edit:
As you can see here that a tensor which is moved to GPU is actually a tensor of type: torch.cuda.*Tensor i.e. torch.cuda.FloatTensor.
So cpu_tensor.to(device) or torch.Tensor([1., 2.], device='cuda') will actually return a tensor of type torch.cuda.FloatTensor.
In which scenario is torch.cuda.Tensor() necessary?
When you want to use GPU acceleration (which is much faster in most cases) for your program, you need to use torch.cuda.Tensor, but you have to make sure that ALL tensors you are using are CUDA Tensors, mixing is not possible here.