Failed to compute dot product of torch.cuda.FloatTensor

1

I used a GPU to compute the dot product of the output of neural networks and a torch.cuda.FloatTensor (both of them are stored in GPU) but got an error saying: TypeError: dot received an invalid combination of arguments - got (torch.cuda.FloatTensor) but expected (torch.FloatTensor tensor).

the codes are like
p = torch.exp(vector.dot(ht))
here vector is a torch FloatTensor and ht is the output of neural networks.

I’ve struggled with these things for days but still got no idea. Thanks in advance for any possible solution!

2

it is saying that one of either vector or ht is not a cuda.FloatTensor but is a FloatTensor. Find it and call .cuda() on it.

1 Like
3

Yes that’s it. Thanks!

4

Hello, Im having a similar issue, im trying to run in gpu a torch.dot with the following code;

loss
loss899×174 20.8 KB

but im obtaining the error;

RuntimeError: _th_dot is not implemented for type torch.cuda.LongTensor

anyone knows what should I modify to make it work ?

Thanks.

5

Try to cast both inputs to float32:

embedding_output = embedding_output.float()
embedding_target =  embedding_target.float()
cos = torch.dot(...)
6

Thank you ! it worked !