Remove device_memory_resource inheritance from all resources and adaptors by bdice · Pull Request #2301 · rapidsai/rmm

bdice · 2026-03-14T13:38:48Z

Summary

Remove device_memory_resource inheritance from all memory resources (stateless, stateful, and adaptors)
Remove do_allocate / do_deallocate / do_is_equal virtual overrides from all resources
Rewrite benchmark factory functions from shared_ptr<device_memory_resource> to any_device_resource
Convert simulated_memory_resource from DMR inheritance to CCCL concepts
Change copy/move from = delete to = default on cuda_async_memory_resource, cuda_async_managed_memory_resource, sam_headroom_memory_resource, and simulated_memory_resource (required for CCCL resource_ref copyability via shared_resource base)
Remove NullUpstream tests and DEVICE_MEMORY_RESOURCE_VIEW_TEST (no longer needed without DMR)

Closes #2295
Part of #2011

copy-pr-bot · 2026-03-14T13:38:52Z

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

@vyasr

…async_resource_ref (#2300) ## Summary Replaces `shared_ptr[device_memory_resource]` with per-subclass `unique_ptr[ConcreteType]` (owning) and `optional[device_async_resource_ref]` (non-owning reference) across all Python/Cython bindings. This is a part of #2011. There are **significant** opportunities to make this Cython code better over time but I have to get something that removes `device_memory_resource` from the Python/Cython side before I can finish migration on the C++ side (#2296). I welcome critique of this design, and ideas for how it can be improved, particularly from @vyasr @wence-. I would like to address any suggested improvements in follow-up PRs, because this changeset is necessary to unblock #2301. The changes in `cdef class DeviceMemoryResource` are perhaps the most significant changes here from a design perspective. The solution I'm going with for now is to keep the `DeviceMemoryResource` class around, as a base class for the Cython MRs, and let it handle allocate/deallocate. It owns a `optional[device_async_resource_ref]` which is used for allocation/deallocation. It's `optional` so that the class can be default-constructed (Cython requires nullary constructors), but it should never be `nullopt` except during initialization. Then, each MR class owns a `c_obj` like `unique_ptr[cuda_memory_resource]`. This is `unique_ptr` so it can be default-constructed for Cython's requirements. I chose `unique_ptr` over `optional` here to emphasize that this member is the thing that actually owns the resource. As with the `c_ref`, this should never be `nullptr` except during initialization. When an MR class is created, it initializes its `c_obj` and then constructs a `c_ref` (a member inherited from the `DeviceMemoryResource` base class). "Special" methods for an MR like getting the statistics counts go through `deref(self.c_obj)`, and "common" methods like allocate/deallocate go through `self.c_ref.value()`. ### Changes - **`.pxd` declarations**: Remove `device_memory_resource` class. Declare `device_async_resource_ref` and a `make_device_async_resource_ref()` inline C++ template that returns `optional` to work around Cython generating default-constructed temporaries for non-default-constructible types. All adaptor constructors take `device_async_resource_ref` instead of `device_memory_resource*`. - **`.pxd` class definitions**: `DeviceMemoryResource` base holds `optional[device_async_resource_ref] c_ref`; each concrete subclass holds `unique_ptr[ConcreteType] c_obj`. - **`.pyx` implementations**: All `__cinit__` methods construct via `unique_ptr` then set `c_ref` via `make_device_async_resource_ref`. Typed accessors (`pool_size`, `flush`, etc.) use `deref(self.c_obj)`. Per-device functions use `set_per_device_resource_ref`. - **`device_buffer.pyx`**: Passes `self.mr.c_ref.value()` instead of `self.mr.get_mr()`. Closes #2294

…tors Remove the device_memory_resource virtual base class inheritance from all production memory resources, adaptors, and stream_ordered_memory_resource. Resources now derive publicly from cuda::mr::shared_resource<Impl> (for stateful/adaptor types) or stand alone with direct CCCL concept methods (for stateless types). The legacy do_allocate/do_deallocate/do_is_equal virtual overrides and pointer-based per-device-resource APIs are removed. stream_ordered_memory_resource provides allocate/deallocate/allocate_sync/ deallocate_sync directly instead of through the DMR virtual dispatch. All 103 C++ tests and 1165 Python tests pass.

bdice · 2026-03-18T00:04:30Z

cpp/include/rmm/detail/cccl_adaptors.hpp

                                                        ::rmm::detail::cccl_resource_ref> and
                             not is_specialization_of_v<std::remove_cv_t<OtherResourceType>,
                                                        ::rmm::detail::cccl_async_resource_ref> and
+                             not shared_resource_cast<OtherResourceType>::value and


This isn't something we want to keep. I don't know how to work around the constraint satisfaction issues in NVIDIA/cccl#8037.

Minimal repro: https://godbolt.org/z/ddfn1sE8Y

Reviewers: I discussed this issue with @ericniebler. We are going to try NVIDIA/cccl#8121 as a workaround. For now, I want to get the current series of open PRs merged (#2301, #2324, #2325), and then try to reduce the workarounds that have been introduced.

bdice · 2026-03-18T00:19:45Z

cpp/tests/mr/aligned_mr_tests.cpp

      void* ptr{nullptr};
      if (i != 0) { std::this_thread::sleep_for(std::chrono::milliseconds{100}); }
-      EXPECT_NO_THROW(ptr = mr.allocate(stream, allocation_size));
+      EXPECT_NO_THROW(ptr = mr.allocate(stream, allocation_size, rmm::CUDA_ALLOCATION_ALIGNMENT));


An explicit alignment argument is required because of the issues described in NVIDIA/cccl#8063. There is no 2-argument allocate(stream, size) overload for shared_resource<T> like there is for other CCCL MR classes. There is only allocation(stream, size, alignment) for shared_resource<T>, and it has no default alignment value.

cpp/benchmarks/random_allocations/random_allocations.cpp

cpp/include/rmm/mr/detail/stream_ordered_memory_resource.hpp

Rewrite benchmark factory functions from shared_ptr<device_memory_resource> to any_device_resource, convert simulated_memory_resource from DMR inheritance to CCCL concepts, and change copy/move from = delete to = default on cuda_async_memory_resource, cuda_async_managed_memory_resource, sam_headroom_memory_resource, and simulated_memory_resource to satisfy CCCL resource_ref copyability requirements.

GCC 13 incorrectly resolves the injected class name to the fully-specialized type inside a class template body, even in template-template parameter position. This causes a hard error (not SFINAE) in is_specialization_of_v when checking against cccl_resource_ref and cccl_async_resource_ref in their own generic-resource constructors. Replace is_specialization_of_v<T, cccl_*_resource_ref> with dedicated is_cccl_*_resource_ref_v<T> traits defined outside the class bodies, where the forward declarations are unambiguously templates. This is correct on both GCC 13 and 14.

github-project-automation bot added this to RMM Project Board Mar 14, 2026

bdice mentioned this pull request Mar 16, 2026

[FEA] Support memory resources from CCCL 3.2 #2011

Open

39 tasks

bdice force-pushed the remove-dmr-inheritance branch from 032b3a6 to 856395e Compare March 17, 2026 08:16

bdice mentioned this pull request Mar 17, 2026

Migrate Python/Cython bindings from device_memory_resource to device_async_resource_ref #2300

Merged

bdice force-pushed the remove-dmr-inheritance branch from 856395e to a06396a Compare March 17, 2026 23:46

bdice moved this to In Progress in RMM Project Board Mar 18, 2026

bdice self-assigned this Mar 18, 2026

bdice added breaking Breaking change improvement Improvement / enhancement to an existing function labels Mar 18, 2026

bdice marked this pull request as ready for review March 18, 2026 00:01

bdice requested review from a team as code owners March 18, 2026 00:01

bdice requested review from ttnghia and vyasr and removed request for a team March 18, 2026 00:01

bdice commented Mar 18, 2026

View reviewed changes

Remove NullUpstream tests

db2eb7c

bdice commented Mar 18, 2026

View reviewed changes

Remove DEVICE_MEMORY_RESOURCE_VIEW_TEST

afd4cc6

ttnghia reviewed Mar 19, 2026

View reviewed changes

cpp/benchmarks/random_allocations/random_allocations.cpp Outdated Show resolved Hide resolved

bdice mentioned this pull request Mar 19, 2026

Ambiguous get_property overload with CCCL's dynamic_accessibility_property #2322

Open

ttnghia reviewed Mar 19, 2026

View reviewed changes

cpp/include/rmm/mr/detail/stream_ordered_memory_resource.hpp Outdated Show resolved Hide resolved

bdice requested a review from a team as a code owner March 19, 2026 18:11

bdice force-pushed the remove-dmr-inheritance branch from 563d7d3 to 8385650 Compare March 20, 2026 20:52

bdice mentioned this pull request Mar 20, 2026

Remove bridge infrastructure and device_memory_resource #2324

Draft

bdice added 2 commits March 20, 2026 16:08

Use == 0 for unsigned types

e4991dc

Use implicit conversion to any_resource in benchmark factory functions

78f58d0

bdice mentioned this pull request Mar 20, 2026

[FEA] Remove device_memory_resource inheritance from all C++ memory resources #2295

Open

22 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Remove device_memory_resource inheritance from all resources and adaptors#2301

Remove device_memory_resource inheritance from all resources and adaptors#2301
bdice wants to merge 7 commits intorapidsai:stagingfrom
bdice:remove-dmr-inheritance

bdice commented Mar 14, 2026 •

edited

Loading

Uh oh!

copy-pr-bot bot commented Mar 14, 2026

Uh oh!

bdice Mar 18, 2026 •

edited

Loading

Uh oh!

bdice Mar 20, 2026

Uh oh!

bdice Mar 18, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

bdice commented Mar 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Uh oh!

copy-pr-bot bot commented Mar 14, 2026

Uh oh!

bdice Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

bdice Mar 20, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

bdice commented Mar 14, 2026 •

edited

Loading

bdice Mar 18, 2026 •

edited

Loading

bdice Mar 18, 2026 •

edited

Loading