dlpack icon indicating copy to clipboard operation
dlpack copied to clipboard
verified publisher icon dmlc

Tag an official v1.0 release

Open leofang opened this issue 1 year ago • 8 comments

Hi @tqchen, we've discussed this in a recent array-API meeting and no objection was raised. Given that several projects have successfully migrated to DLPack v1.0, including:

  • NumPy (https://github.com/numpy/numpy/pull/26501)
  • mpi4py (https://github.com/mpi4py/mpi4py/pull/494)
  • CUDA Python (https://github.com/NVIDIA/cuda-python/pull/87)

and the release candidate has been cut several months ago (in February), it should be no issue in actually tagging a v1.0 release 🙂 Thanks!

leofang avatar Sep 09 '24 14:09 leofang

cc: @seberg @rgommers @oleksandr-pavlyk for vis

leofang avatar Sep 09 '24 14:09 leofang

Just did that, we should keep track of migration status of other projects, including pytorch

tqchen avatar Sep 09 '24 15:09 tqchen

Thanks, Tianqi!

we should keep track of migration status of other projects, including pytorch

Totally agreed. I've shared the DLPack update news internally a few months back. Will try to follow up with more projects.

leofang avatar Sep 09 '24 17:09 leofang

  • dpctl.tensor also supports DLPack 1.0 (doc), but the docstring does not reflect that at the moment.

oleksandr-pavlyk avatar Oct 17 '24 17:10 oleksandr-pavlyk

  • CuPy PR for 1.0 support: https://github.com/cupy/cupy/pull/8683
  • PyTorch PR should be in the works
  • Cc @jakevdp for JAX.

rgommers avatar Oct 31 '24 20:10 rgommers

It's difficult to figure out whether we're conforming to the spec or not (there's no conformance test suite, right?), but I believe the only thing JAX is missing is using DLManagedTensorVersioned rather than DLManagedTensor. Guess I'll do that.

hawkinsp avatar Oct 31 '24 21:10 hawkinsp

It's difficult to figure out whether we're conforming to the spec or not (there's no conformance test suite, right?),

I think there's no library-agnostic test suite, no. It'd be quite nice to have that - there's a TODO in https://github.com/data-apis/array-api-tests/issues/301 for it, and that's probably the right place at least for Python-level tests - and that test suite is already run in JAX, so it'd be easy logistically.

Perhaps borrowing from NumPy (https://github.com/numpy/numpy/blob/main/numpy/_core/tests/test_dlpack.py) or some other implementation is helpful for now..

rgommers avatar Feb 04 '25 16:02 rgommers

PyTorch PR with implementation in the works: https://github.com/pytorch/pytorch/pull/145000

rgommers avatar Feb 04 '25 16:02 rgommers

PyTorch 2.9.0 was just released, and included the upgrade to DLPack 1.0

rgommers avatar Oct 15 '25 17:10 rgommers