CuPy to PyTorch with __cuda_array_interface__ - Array Strides not Multiple of Element Byte Size

PyTorch version: torch.__version__: 1.3.0

CuPy config:

CuPy Version          : 7.0.0rc1
CUDA Root             : /usr/local/cuda-10.0
CUDA Build Version    : 10000
CUDA Driver Version   : 10020
CUDA Runtime Version  : 10000
cuDNN Build Version   : None
cuDNN Version         : None
NCCL Build Version    : None
NCCL Runtime Version  : None

With PyTorch accepting the cuda_array_interface, I’d expect to be able to use a CuPy generated CUDA array in PyTorch without going through DLPack.

import cupy as cp
import torch

a = cp.random.rand(10000)
b = torch.as_tensor(a)

Throws: ValueError: given array strides not a multiple of the element byte size. Make a copy of the array to reallocate the memory.

If this is on PyTorch’s side, please let me know, and I’ll file an issue there. Thanks!

[jakirkham]

Also just a small note, using the following code (based off the code in the OP), we notice something peculiar

import cupy as cp
import torch

a = cp.random.rand(10000)
b = torch.as_tensor(a)
assert b.is_cuda  # fails

This happens because PyTorch has actually done a device-to-host transfer under-the-hood. To fix this we have to specify the device.

import cupy as cp
import torch

a = cp.random.rand(10000)
b = torch.as_tensor(a, device=torch.device('cuda'))
assert b.is_cuda  # works!

One can also inspect __cuda_array_interface__ of both objects to confirm that this not only remains on device, but was a zero-copy conversion. Thus b uses the same memory as a.

[awthomp]

Confirmed fixed with https://github.com/pytorch/pytorch/pull/24947