Use CCCL cuda::mr resource aliases instead of RMM aliases

[bdice]

Summary

RAPIDS libraries name memory resources in their public APIs through RMM aliases such as rmm::device_async_resource_ref. These aliases are thin names over CCCL types (cuda::mr::resource_ref<...>), but spelling them through RMM forces every public header to include <rmm/resource_ref.hpp> and keeps RMM in the API surface of every library.

CCCL now provides named aliases in cuda::mr (NVIDIA/cccl#11751, available in CCCL 3.6 via rapidsai/rapids-cmake#1097). Every library should use these aliases directly, with no per-library aliases. This removes RMM headers from public APIs and gives all libraries one shared spelling, while leaving the mangled symbols unchanged, since an alias is not part of the ABI.

This is part of #321 ("Use CCCL memory resources instead of rmm memory resources") and rapidsai/rmm#2455.

Aliases

Replace	With	CCCL type
rmm::device_async_resource_ref	cuda::mr::device_resource_ref	cuda::mr::resource_ref<cuda::mr::device_accessible>
rmm::host_async_resource_ref	cuda::mr::host_resource_ref	cuda::mr::resource_ref<cuda::mr::host_accessible>
rmm::host_device_async_resource_ref	cuda::mr::host_device_resource_ref	cuda::mr::resource_ref<cuda::mr::host_accessible, cuda::mr::device_accessible>
cuda::mr::any_resource<cuda::mr::device_accessible>	cuda::mr::any_device_resource
cuda::mr::any_resource<cuda::mr::host_accessible>	cuda::mr::any_host_resource
cuda::mr::any_resource<cuda::mr::host_accessible, cuda::mr::device_accessible>	cuda::mr::any_host_device_resource

Plain names are stream-ordered, matching CCCL where resource_ref is stream-ordered and synchronous_resource_ref is the explicit exception. CCCL does not provide synchronous aliases. Synchronous refs are spelled as cuda::mr::synchronous_resource_ref<...>, except for RAFT's host-only PMR/mmap resources, which keep raft::mr::synchronous_host_resource_ref.

RAFT migrates its pinned/managed resource APIs to cuda::mr::host_device_resource_ref and cuda::mr::any_host_device_resource. Its sync-only CCCL legacy defaults are wrapped in cuda::mr::synchronous_resource_adapter. Existing host-device aliases are deprecated and now name the stream-ordered types; custom sync-only resources require an adapter. The host-only raft::mr::host_resource_ref remains synchronous under the explicit raft::mr::synchronous_host_resource_ref name, with the old name deprecated.

RMM uses the CCCL aliases internally, and deprecates and removes its own aliases (rapidsai/rmm#2562).

Migration per library

Each library does the following in a single PR:

• Replace rmm::*resource_ref and cuda::mr::any_resource<...> uses with the cuda::mr:: aliases in C++ (public headers, sources, tests, benchmarks, examples).
• Replace #include <rmm/resource_ref.hpp> with #include <cuda/memory_resource>.
• Where the library has Cython bindings, stop importing device_async_resource_ref from rmm.librmm.memory_resource, and carry a library-owned Cython declaration of the ref (a copy of RMM's default-constructible wrapper, retargeted at cuda::mr::device_resource_ref).

Validation: the exported symbols of the library's shared object are identical before and after.

Pull requests

rmm::*resource_ref uses in each repository (approximate, from main at the time of writing):

• cuDF: ~3200 C++ uses, ~100 Cython files. Reference implementation for the other libraries. Use CCCL memory resource aliases directly NVIDIA/cudf#24316
• RAFT: ~40 C++ uses. Use stream-ordered pinned/managed resource APIs; retain the explicit synchronous alias only for host-only PMR/mmap resources. Use CCCL aliases and stream-ordered pinned and managed resources NVIDIA/raft#3159
• cuVS: ~50 C++ uses. Use CCCL memory resource aliases directly NVIDIA/cuvs#2699
• cuGraph: ~30 C++ uses. Use CCCL memory resource aliases directly cugraph#5685
• RapidsMPF: ~20 C++ uses, 1 Cython file Use CCCL memory resource aliases directly rapidsmpf#1238
• cuML: ~6 C++ uses. Use CCCL memory resource aliases directly NVIDIA/cuml#8715
• cuGraph-GNN: ~3 C++ uses, all in one test file. Pass statistics resources directly instead of using RMM resource ref aliases cugraph-gnn#548
• RMM: use CCCL aliases internally, deprecate the RMM aliases after the libraries above have migrated, and remove them in 27.02 ([FEA] Deprecate and remove resource ref aliases in favor of CCCL cuda::mr aliases rmm#2562)
  • Use CCCL aliases internally: Use CCCL memory resource aliases directly rmm#2563
  • Deprecate the aliases
  • Remove the aliases in 27.02