Repository navigation
[FEA]: Support importing __c_dlpack_exchange_api__ objects as StridedMemoryView #1439
Description
Description
Add support for importing objects that implement the __c_dlpack_exchange_api__ protocol as StridedMemoryView objects.
Motivation
The __c_dlpack_exchange_api__ protocol provides a C-level interface for efficient zero-copy data exchange. Currently, StridedMemoryView supports construction from Python-level protocols (__dlpack__, __cuda_array_interface__, buffer protocol), but does not support the C-level exchange API.
Supporting this protocol would enable:
- More efficient data sharing with libraries that expose the C-level API
- Reduced overhead by avoiding Python-level protocol dispatch
- Better integration with Cython-based libraries that use the C exchange API
Requested Features
- Add a class method or constructor path in
StridedMemoryViewthat can consume objects implementing__c_dlpack_exchange_api__ - Properly handle the DLManagedTensor lifecycle (deleter callback)
Additional Context
- Related to [FEA]: Support exporting
StridedMemoryViewobjects via DLPack #1438 (export support) - DLPack spec: https://dmlc.github.io/dlpack/latest/
Activity
leofang commentedon May 8, 2026
TBH I am not convinced this has any real benefit, mainly because this is a very new and somewhat tailored solution that not many libraries have adopted (ex: we do not offer this in CuPy); for PyTorch, starting v1.0.0
torch.Tensorgoes through the tensor bridge (#1894) so this is moot anyway.In any case, the work was done in #1630 and was shipped in v0.7.0. Closing.