Support for torch.mm with Sparse Half Tensors? "addmm_sparse_cuda" not implemented for Half

[sbonner0]

Hi,

I am trying to perform sparse and dense matrix multiplication using half precision tensors in pytorch.

The following code:

import torch
a = torch.randn(3,2).half().cuda()
i = torch.LongTensor([[0, 1, 1],  [2, 0, 2]]) 
v = torch.FloatTensor([3, 4, 5]) 
b = torch.sparse.FloatTensor(i, v, torch.Size([2,3])).half().cuda()
c = torch.spmm(b, a)

will produce this error: RuntimeError: "addmm_sparse_cuda" not implemented for 'Half'

Is there anyway to solve this?

I am unsure if this is a Apex or Pytorch issue so I thought I would post it in both places.

[sbonner0]

This is an issue with Pytorch and not Apex itself - see this issue: pytorch/pytorch#41069