__cuda_array_interface__ conversion does not support readonly arrays

[hawkinsp]

PyTorch does not support importing readonly GPU arrays via __cuda_array_interface__:
https://numba.pydata.org/numba-doc/dev/cuda/cuda_array_interface.html

pytorch/torch/csrc/utils/tensor_numpy.cpp

Line 292 in 2471ddc

throw TypeError("the read only flag is not supported, should always be False");

To Reproduce

If you have a copy of jax and jaxlib with GPU support built from head, the following will reproduce:

In [1]: import torch, jax, jax.numpy as jnp

In [2]: x = jnp.array([1,2,3])

In [3]: y = torch.as_tensor(x)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-3-f43755a1deef> in <module>
----> 1 y = torch.tensor(x)

TypeError: the read only flag is not supported, should always be False

(Aside: this is not a PyTorch bug, but curiously CuPy drops the readonly flag, so you can make the import "work" by laundering the array through CuPy:

In [1]: import torch, jax, jax.numpy as jnp, cupy

In [2]: x = jnp.array([1,2,3])

In [3]: y = torch.as_tensor(cupy.asarray(x), device="cuda")

In [4]: x.__cuda_array_interface__
Out[4]:
{'shape': (3,),
 'typestr': '<i4',
 'data': (140492215944704, True),
 'version': 2}

In [5]: y.__cuda_array_interface__
Out[5]:
{'typestr': '<i4',
 'shape': (3,),
 'strides': (4,),
 'data': (140492215944704, False),
 'version': 1}
)

Expected behavior

PyTorch should support the readonly flag.

cc @ngimel

[michaelklachko]

Sorry for off-topic, but I'm curious about y = torch.as_tensor(cupy.asarray(x), device="cuda") line - I was looking for a way to pass torch cuda tensors to cupy for some ops not available in pytorch, then passing them back to pytorch all while keeping the data on the same GPU. I thought the best way to do this is to use DLPack (e.g. see this cupy page), but your example seems to indicate that DLPack is unnecessary. Is there any reason to use DLPack in this scenario?