EPIC: Port Rapids projects from pynvml to cuda.core.system

This is an EPIC to track porting usage of pynvml to cuda.core.system across the RAPIDS family of projects. In order of importance / level of NVML usage.

Naming fixes prereq: #1946

dask_cuda
- cuda.core.system: Add MIG-related APIs #1916
- Migrate from pynvml to cuda.core.system rapidsai/dask-cuda#1645
cudf
jupyterlab-nvdashboard
rapids-cli
- cuda.core.system: Add basic Nvlink and Utilization support #1918
- Migrate pynvml to cuda.core.system rapidsai/rapids-cli#146
deployment
- Mention cuda.core.system as an option for getting NVML information. rapidsai/deployment#677
ucxx
- cuda.core.system: Add MIG-related APIs #1916
- Migrate pynvml to cuda.core.system rapidsai/ucxx#640
auto-nvidia-accelerator (This should have been archived)
- https://github.com/rapidsai/auto-nvidia-accelerator/pull/182
rapidsmpf
- Migrate from pynvml to cuda.core.system rapidsai/rapidsmpf#969
cuml
- Migrate pynvml.py to cuda.core.system cuml#8022
cugraph
- Migrate pynvml to cuda.core.system for GPU metrics polling rapidsai/cugraph#5501
kvikio
- Migrate from pynvml.py to cuda.core.system rapidsai/kvikio#955
rapids-examples
- Migrate from pynvml.py to cuda.core.system rapidsai/rapids-examples#63
tpm-tools
- Seems like examples? Not entirely sure. Deferring.

Migrate from pynvml to cuda.core.system (#969)

Thanks Mike! 🙏

In many cases we run the benchmarks and examples on CI. So we may still want to migrate them if the goal is dropping PyNVML

That said, think the prioritization here is reasonable. Also we are very appreciative on the RAPIDS side of you picking this up and pushing it forward

Merging many of these Rapids PRs above is dependent on changing how dependencies are specified: rapidsai/build-planning#285

@mdboom

Merging many of these Rapids PRs above is dependent on changing how dependencies are specified: rapidsai/build-planning#285

I've just finished rapidsai/build-planning#285. Across RAPIDS, we no longer depend directly on cuda-python in any libraries (even in testing environments), and instead have these:

cuda-bindings >=12.9.2,<13.0  # (CUDA 12)
cuda-bindings >=13.0.1,<14.0  # (CUDA 13)
cuda-core>=0.5.1

Most projects only depend on cuda-bindings, cuda-core is only used in a handful of libraries today.

I went with those ranges because they most closely matched what cuda-python was pulling in for us. @mdboom looks like your series of PRs is proposing moving those floors higher, we'd appreciate your input on why: NVIDIA/cudf#22305 (comment)

@jameslamb

@jameslamb I would recommend tightening the cuda-core pin if possible. There was a number of breaking changes introduce in cuda-core 1.0, after which we are supporting strict semantic versioning until we release a 2.0 which would be aligned with CUDA 14. At the very least would recommend pinning to <2.0 as well.

@kkraus14

Agree with @kkraus14 on moving to the API-stable cuda-core 1.0.0.

Additionally, NVML support wasn't added to cuda-bindings until 13.2.0 (and 12.9.6 on the 12.x series).

I would like to transfer this issue to our private repo, but it's in a different org now 🤔

@mdboom

@mdboom @kkraus14 we're no longer seeing major issues here and we just bumped our cuda-core requirement to 1.0 across RAPIDS. The main outstanding concern is that this forces us to require numba-cuda 0.30.2, but since we're actively working on our migration to numba-cuda-mlir my hope is that by our next release (26.12) we will no longer carry a numba-cuda dependency and so requiring such a new version of it will not cause any issues.

@mdboom would you be able to revive your pynvml PRs now? Apologies that we took so long to unblock you.

Navigation Menu

Repository navigation

EPIC: Port Rapids projects from pynvml to cuda.core.system #1915

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

EPIC: Port Rapids projects from pynvml to cuda.core.system #1915

Description

Activity

jakirkham commented on Apr 23, 2026

mdboom commented on Jun 10, 2026

jameslamb commented on Jul 7, 2026

kkraus14 commented on Jul 8, 2026

mdboom commented on Jul 14, 2026

leofang commented on Sep 2, 2026

vyasr commented on Oct 13, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions