Add numba-cuda-mlir docs, cuda-core dependency, and toolkit extras by leofang · Pull Request #2101 · NVIDIA/cuda-python

leofang · 2026-05-18T14:47:52Z

Summary

Mention numba-cuda-mlir in the CUDA Python README and metapackage docs.
Add cuda-core~=1.0.0 as a required dependency of the cuda-python metapackage.
Add CUDA Toolkit metapackage dependencies to cuda-bindings[all] and cuda-core's cu12 / cu13 extras for improved dependency resolution.
Keep cuDLA available through cuda-bindings[all] on Linux aarch64 via a guarded nvidia-cudla==13.* dependency.
Update installation docs to include nvFatBin and cuDLA in the optional dependency list.

Notes

cuda-cccl is intentionally left out of the cuda-python metapackage for now because it does not yet have Python 3.14 free-threaded builds. It can be added in a follow-up once that support is available.

… cuda-toolkit extras - Mention numba-cuda-mlir in README.md and DESCRIPTION.rst above the existing numba.cuda entry - Add cuda-core~=1.0.0 and cuda-cccl~=1.0.0 as required dependencies of the cuda-python metapackage (closes NVIDIA#148, closes NVIDIA#691) - Add bare cuda-toolkit==13.* to cuda-bindings[all] and remove cudla from the component-specific extras (closes NVIDIA#903) - Add cuda-toolkit==12.*/13.* to cuda-core's cu12/cu13 extras

copy-pr-bot · 2026-05-18T14:47:56Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

leofang · 2026-05-18T14:57:05Z

 [project.optional-dependencies]
 all = [
-    "cuda-toolkit[nvrtc,nvjitlink,nvvm,nvfatbin,cudla]==13.*",
+    "cuda-toolkit[nvrtc,nvjitlink,nvvm,nvfatbin]==13.*",


I noticed this change sneaked in from PR #2034. This is a bit problematic because cuDLA is not a universally available component in the CTK, and the majority of users cannot use it but they'd see this at install time:

WARNING: cuda-toolkit 13.1.2.0 does not provide the extra 'cudla'

We should just document that DLA users need to install cuDLA if they haven't already (which I suspect is not the case, because there should be a system CTK already installed on embedded devices).

I approved this knowingly, after discovering that the meta package has the filters we need.

I missed that warning message though. If we want to get rid of it, we have do add the conditions similar to what we're doing under cuda_pathfinder/pyproject.toml:

"nvidia-cudla; platform_system == 'Linux' and platform_machine == 'aarch64'",

That would still fire on aarch64 server machines where I don't think we have a way to differentiate Jetson / Tegra with current pip capabilities today

My 2c: I think we should include cuDLA in all still as it would only error at runtime if someone tried to use a cuDLA binding from a non Tegra system. The cuDLA binary is tiny (< 1MB), so binary size impact isn't a big concern.

@kkraus14 My concerns:

the warning (see above)

would cuda-toolkit==13.0 still work for constraining the installed packages? cuDLA wheel was not shipped until 13.2

If we guard it based on @rwgk's comment in using platform_system and platform_machine does it still warn?

I assume if we added it to all and someone did pip install cuda-bindings[all] cuda-toolkit==13.0 that it would error because there isn't a cuda-toolkit[cudla] package for 13.0?

Done: commit f4c6932

leofang · 2026-05-22T18:40:41Z

/ok to test 0587448

leofang · 2026-05-22T20:42:45Z

    install_requires=[
        f"cuda-bindings{matcher}{version}",
+        "cuda-core~=1.0.0",
+        "cuda-cccl~=1.0.0",


Life has been not very kind to us. Because cuda-cccl does not yet have a 3.14t build, and because there is no environment marker that would allow us to append a condition here (PEP 780 is still in draft status and the DPO discussion is brutal), all Python 3.14t pipelines failed. We have no other way but drop cuda-cccl here:

Suggested change

"cuda-cccl~=1.0.0",

Let's add cuda-cccl in a follow up because of this.

Done: commit f4c6932

@shwina for visibility

kkraus14 · 2026-05-23T01:42:18Z

 * Python 3.10 - 3.14
 * Driver: Linux (580.65.06 or later) Windows (580.88 or later)
-* Optionally, NVRTC, nvJitLink, NVVM, and cuFile from CUDA Toolkit 13.x
+* Optionally, NVRTC, nvJitLink, nvFatBin, NVVM, and cuFile from CUDA Toolkit 13.x


Should we include cuDLA here now as well?

Done: commit f4c6932

@nikshayshrivastava for visibility

kkraus14 · 2026-05-23T01:42:49Z


 * ``nvidia-cuda-nvrtc`` (NVRTC runtime compilation library)
 * ``nvidia-nvjitlink`` (nvJitLink library)
+* ``nvidia-nvfatbin`` (nvFatBin library)


+1 cuDLA here?

Done: commit f4c6932

@nikshayshrivastava for visibility

kkraus14 · 2026-05-23T01:46:21Z

 [project.optional-dependencies]
 all = [
-    "cuda-toolkit[nvrtc,nvjitlink,nvvm,nvfatbin,cudla]==13.*",
+    "cuda-toolkit[nvrtc,nvjitlink,nvvm,nvfatbin]==13.*",


My 2c: I think we should include cuDLA in all still as it would only error at runtime if someone tried to use a cuDLA binding from a non Tegra system. The cuDLA binary is tiny (< 1MB), so binary size impact isn't a big concern.

kkraus14 · 2026-05-23T01:47:24Z

    install_requires=[
        f"cuda-bindings{matcher}{version}",
+        "cuda-core~=1.0.0",
+        "cuda-cccl~=1.0.0",


Let's add cuda-cccl in a follow up because of this.

Drop cuda-cccl from the metapackage until free-threaded Python wheels are available, and keep cuDLA in cuda-bindings[all] via a guarded direct dependency.

Keep cuda-bindings[all] consistent with the CUDA 13 toolkit dependencies when installing cuDLA on Linux aarch64.

rwgk

Looks good to me.

I looked through manually, and I used GPT-5.5 for a full agent review.

rwgk · 2026-05-25T19:03:03Z

/ok to test 6b8d099

github-actions · 2026-05-25T21:48:01Z

Doc Preview CI
Preview removed because the pull request was closed or merged.

github-actions Bot added cuda.bindings Everything related to the cuda.bindings module cuda.core Everything related to the cuda.core module labels May 18, 2026

Add numba-cuda-mlir to cuda-python docs index.rst

9dcb465

leofang commented May 18, 2026

View reviewed changes

leofang added documentation Improvements or additions to documentation P0 High priority - Must do! packaging Anything related to wheels or Conda packages labels May 18, 2026

leofang added this to the cuda.bindings 13.3.0 & 12.9.7 milestone May 18, 2026

leofang self-assigned this May 19, 2026

leofang commented May 19, 2026

View reviewed changes

Comment thread cuda_python/setup.py Outdated

leofang requested review from aterrel, danielfrg and kkraus14 May 19, 2026 21:57

leofang added 2 commits May 22, 2026 01:05

Update install.rst

06afca8

Merge branch 'main' into meta-deps

0587448

leofang marked this pull request as ready for review May 22, 2026 05:07

gmarkall approved these changes May 22, 2026

View reviewed changes

leofang enabled auto-merge (squash) May 22, 2026 18:41

rparolin approved these changes May 22, 2026

View reviewed changes

This comment has been minimized.

Sign in to view

leofang commented May 22, 2026

View reviewed changes

kkraus14 reviewed May 23, 2026

View reviewed changes

rwgk added 2 commits May 25, 2026 11:11

Merge branch 'main' into leofang→meta-deps

ebd2109

Adjust PR 2101 dependencies for 13.3

f4c6932

Drop cuda-cccl from the metapackage until free-threaded Python wheels are available, and keep cuDLA in cuda-bindings[all] via a guarded direct dependency.

rwgk marked this pull request as draft May 25, 2026 18:42

auto-merge was automatically disabled May 25, 2026 18:42
Pull request was converted to draft

Constrain cuDLA optional dependency to CUDA 13

6b8d099

Keep cuda-bindings[all] consistent with the CUDA 13 toolkit dependencies when installing cuDLA on Linux aarch64.

rwgk changed the title ~~Add numba-cuda-mlir docs, cuda-core/cuda-cccl deps, cuda-toolkit extras~~ Add numba-cuda-mlir docs, cuda-core dependency, and toolkit extras May 25, 2026

rwgk approved these changes May 25, 2026

View reviewed changes

rwgk marked this pull request as ready for review May 25, 2026 19:02

rwgk merged commit 6959cb9 into NVIDIA:main May 25, 2026
100 checks passed

rwgk mentioned this pull request May 25, 2026

Add cuda-cccl as a required dependency to cuda-python #691

Open

Conversation

leofang commented May 18, 2026 • edited by rwgk Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Notes

Uh oh!

copy-pr-bot Bot commented May 18, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

leofang commented May 22, 2026

Uh oh!

This comment has been minimized.

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rwgk left a comment

Choose a reason for hiding this comment

Uh oh!

rwgk commented May 25, 2026

Uh oh!

Uh oh!

github-actions Bot commented May 25, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

leofang commented May 18, 2026 •

edited by rwgk

Loading