weird new() method of torch.cuda.sparse tensor #2249
Description
Let say,
a = torch.sparse.FloatTensor(torch.LongTensor([[0]]), torch.FloatTensor([0]), torch.Size([1]))
a = a.cuda(1)
when I try,
i = torch.LongTensor([[0, 1], [2, 0]])
v = torch.FloatTensor([3, 4])
i = i.cuda(1)
v = v.cuda(1)
b = a.new(i, v, torch.Size([2,3]))
it throws error: arguments are located on different GPUs
But when I try,
i = i.cuda(0)
v = v.cuda(0)
b = a.new(i, v, torch.Size([2,3]))
then it will make b a sparse tensor on GPU 0.
Besides,
a = a.new()
print(a.get_device()) # sometimes print 0 and sometimes print -1