RuntimeError: _th_dot not supported on CUDAType for Short

1

Why I cannot perform dot operation on two matrices?


print(S.shape, S.dtype)
print(G.shape, G.dtype)
torch.dot(S, torch.t(G))


torch.Size([256, 31360]) torch.int16
torch.Size([1825, 31360]) torch.int16

RuntimeError: _th_dot not supported on CUDAType for Short

Same happens with int and long.

2

Hi,

Could please sure a code that can reproduces this error?

The problem is torch.dot() only supports 1D tensors. So, the error should be about it. The alternatives function for 2D and 3D (batch-wise) is torch.mm and torch.bmm respectively.
Here is the code I used to produce your case (although could not get your error):

a = torch.randint(high=10, size=(5, 7), dtype=torch.int16)
b = torch.randint(high=10, size=(5, 7), dtype=torch.int16)
torch.dot(a, torch.t(b))

# output 
# RuntimeError: 1D tensors expected, got 2D, 2D tensors at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:431

Maybe you are using an old version of PyTorch?