René's URL Explorer Experiment

{"@context":"https://schema.org","@type":"DiscussionForumPosting","headline":"RFC: Change `__cuda_stream__` from class attribute to method","articleBody":"One of the reasons that people complain about CUDA Array Interface (CAI) being slow is because it has side effects. Consider this example:\n```python\n>>> class KKK:\n...     @property\n...     def __proppp__(self):\n...             print(\"I am called\")\n... \n>>> \n>>> type(KKK.__proppp__)\n\u003cclass 'property'>\n>>> k = KKK()\n>>> hasattr(k, \"__proppp__\")\nI am called\nTrue\n```\nOne can see that the `hasattr` call on any class attribute/property would actually execute the attribute implementation, but not if it is a method:\n```python\n>>> class KKK:\n...     def __proppp__(self):\n...             print(\"I am called\")\n... \n>>> \n>>> type(KKK.__proppp__)\n\u003cclass 'function'>\n>>> k = KKK()\n>>> hasattr(k, \"__proppp__\")\nTrue\n```\nI realized I had forgotten about this during the design stage until now. This fact has also been [pointed out by @gmarkall](https://big-grey.co.uk/2020/02/27/surprises-with-hasattr-and-the-property-decorator/) (back in the days of designing CAI, IIRC).\n\nI propose to change `__cuda_stream__` and any future protocols to a callable class method from a class property/attribute so as to avoid side effects (thereby reducing overhead) and establish a future-proof convention, following DLPack and other protocols.","author":{"url":"https://github.com/leofang","@type":"Person","name":"leofang"},"datePublished":"2025-01-03T20:28:04Z","url":"https://github.com/NVIDIA/cuda-python/issues/348"}