Repository navigation
[FEA]: Support inheritance from Buffer #2435
Description
Is this a duplicate?
- I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct
Area
cuda.core
Is your feature request related to a problem? Please describe.
nvmath-python is considering creating a subclass of Buffer specific to NVSHMEM because we want to mostly follow cuda.core's semantics for Buffer/MemoryResource, but need some different behavior regarding garbage collection.
Describe the solution you'd like
Buffer should use classmethod instead of staticmethod for all methods that return an instance of the class. i.e. from_handle() and _reduce_helper()
Describe alternatives you've considered
Composing Buffer with our Buffer.
Replicating the Buffer protocol with an unrelated class.
Rolling our own MemoryResource/Buffer abstraction that is unique to nvmath's distributed APIs.
Additional context
Related to: #750
Activity
mdboom commentedon Jul 29, 2026
As an aside, do we want to consider a policy of favoring
@classmethodover@staticmethodthroughout CUDA core? I think unless there is a good reason to make a class "final" (in the C++ sense) we shouldn't make it hard to subclass.@Andy-Jost: thoughts?
leofang commentedon Nov 4, 2026
I recall in an internal thread around Aug 13 we discussed that we prefer composition over inheritance, except for memory resources inheriting from the abstract base class. @carterbox is this still needed?
carterbox commentedon Nov 5, 2026
We've decided to take option 3: "Rolling our own MemoryResource/Buffer abstraction that is unique to nvmath's distributed APIs."