Add `get_cuda_native_handle` #773

leofang · 2025-07-22T13:46:13Z

Description

closes #564

The performance is comparable to the existing int() approach (O(10) ns):

In [1]: from cuda.bindings import driver

In [3]: s = driver.CUstream(1234)

In [4]: int(s)
Out[4]: 1234

In [6]: from cuda.bindings.utils import get_cuda_native_handle

In [7]: get_cuda_native_handle(s)
Out[7]: 1234

In [8]: %timeit int(s)
44.7 ns ± 1.34 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

In [9]: %timeit get_cuda_native_handle(s)
95.1 ns ± 1.06 ns per loop (mean ± std. dev. of 7 runs, 10,000,000 loops each)

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

copy-pr-bot · 2025-07-22T13:46:17Z

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 · 2025-07-22T13:47:50Z

(added a short benchmark to the PR description)

leofang · 2025-07-22T13:52:35Z

/ok to test 8970dca

leofang · 2025-07-22T14:23:25Z

/ok to test 4ca551c

leofang · 2025-07-22T14:24:47Z

cuda_bindings/setup.py

@@ -287,6 +288,7 @@ def prep_extensions(sources, libraries):

 # new path for the bindings from cybind
 def rename_architecture_specific_files():
+    path = os.path.join("cuda", "bindings", "_internal")


Note: this is a bug fix exposed by the new line 224. Previously, the path variable happened to be right (leaking from the loop variable at line 227), but it's not longer correct after we append one more item to path_list.

leofang · 2025-07-23T01:40:33Z

Decision to make: Should we raise DeprecationWarning in __int__()? It can be very noisy but it's the only effective way for us to inform users about the deprecation.

rwgk

Decision to make: Should we raise DeprecationWarning in __int__()? It can be very noisy but it's the only effective way for us to inform users about the deprecation.

Do we want to give users a grace period, in case they need to support two cuda-bindings versions for a while?

Concretely:

In the next release, deprecate only in the documentation.

Only with the next minor release bump: actually raise DeprecationWarning.

cuda_bindings/tests/test_utils.py

leofang · 2025-07-23T09:23:17Z

/ok to test 53055e5

rwgk · 2025-07-23T16:32:35Z

Decision to make: Should we raise DeprecationWarning in __int__()? It can be very noisy but it's the only effective way for us to inform users about the deprecation.

Do we want to give users a grace period, in case they need to support two cuda-bindings versions for a while?

Concretely:

In the next release, deprecate only in the documentation.

Only with the next minor release bump: actually raise DeprecationWarning.

@leofang WDYT about introducing a "Phased deprecations" section somewhere, so we don't forget in the future.

E.g. for this case:

__int__() method for type A, B, C were informally deprecated in cuda-bindings release vX.Y.Z. DeprecationWarning will be raised starting with release vX.Y+1.0.

github-actions · 2025-07-23T16:45:37Z

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

leofang · 2025-07-23T17:05:30Z

Yup that's great idea, tracked in #775.

rwgk · 2025-07-23T23:30:37Z

@leofang This PR is causing a weird issue under Windows. I first noticed when updating the unreleased branch, but the issue is actually showing up already in the CI logs here, e.g. this run against main right after this PR was merged:

https://github.com/NVIDIA/cuda-python/actions/runs/16476401578/job/46579899006

All tests pass actually, but each invocation of python generates some noise. Have you seen something like this before?

Running bindinds tests
/c/a/cuda-python/cuda-python/cuda_bindings
Error processing line 4 of C:\a\_tool\Python\3.12.10\x64\Lib\site-packages\_cuda_bindings_redirector.pth:

  Traceback (most recent call last):
    File "<frozen site>", line 206, in addpackage
    File "<string>", line 1, in <module>
    File "C:\a\_tool\Python\3.12.10\x64\Lib\site-packages\_cuda_bindings_redirector.py", line 28, in <module>
      import cuda.bindings
    File "C:\a\_tool\Python\3.12.10\x64\Lib\site-packages\cuda\bindings\__init__.py", line 16, in <module>
      from cuda.bindings import utils
    File "C:\a\_tool\Python\3.12.10\x64\Lib\site-packages\cuda\bindings\utils\__init__.py", line 4, in <module>
      from ._get_handle import get_cuda_native_handle
    File "cuda\\bindings\\utils\\_get_handle.pyx", line 1, in init cuda.bindings.utils._get_handle
    File "cuda\\bindings\\_lib\\utils.pyx", line 1, in init cuda.bindings._lib.utils
    File "cuda\\bindings\\driver.pyx", line 1, in init cuda.bindings.driver
    File "cuda\\bindings\\cydriver.pyx", line 1, in init cuda.bindings.cydriver
    File "cuda\\bindings\\_bindings\\cydriver.pyx", line 6, in init cuda.bindings._bindings.cydriver
  ModuleNotFoundError: No module named 'win32api'

Remainder of file ignored
============================= test session starts =============================
platform win32 -- Python 3.12.10, pytest-8.4.1, pluggy-1.6.0 -- C:\a\_tool\Python\3.12.10\x64\python.exe
cachedir: .pytest_cache
benchmark: 5.1.0 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: C:\a\cuda-python\cuda-python\cuda_bindings\tests
configfile: pytest.ini
plugins: benchmark-5.1.0
collecting ... collected 176 items / 1 skipped

tests\test_cuda.py::test_cuda_memcpy PASSED                              [  0%]
... <snip> ...
================= 171 passed, 6 skipped, 3 warnings in 52.02s =================

leofang · 2025-07-23T23:42:38Z

I've seen errors locally like this too, but I thought it was because I had a corrupted environment and it did go away after a clean rebuild. I probably know what's going on, let me see if I can put together a patch quickly.

rwgk · 2025-07-23T23:46:41Z

it did go away after a clean rebuild

That's surprising. I did multiple clean rebuilds (while backtracking from the unreleased testing), for me the issue was reliably reproducible.

let me see if I can put together a patch quickly.

Sounds great, thanks!

… suggested by ChatGPT) to resolve NVIDIA#773 (comment)

* Simplify cuda_bindings/site-packages/_cuda_bindings_redirector.py (as suggested by ChatGPT) to resolve #773 (comment) * Avoid premature import of `cuda.bindings` in redirector to fix startup error Previously, the `_cuda_bindings_redirector.py` file imported `cuda.bindings` at interpreter startup to ensure the `cuda` namespace was initialized. However, this triggered early loading of Cython modules that depend on `win32api`, which may not be available yet due to `.pth` file processing order. This commit replaces `import cuda.bindings` with `import cuda`, which is sufficient to ensure the `cuda` namespace is in `sys.modules`. The version redirect via `cuda.__version__` is preserved by assigning a custom `LazyCudaModule` type to `sys.modules["cuda"]`. As a result, `cuda.__version__` continues to work (with a deprecation warning), while avoiding startup errors when `win32api` is not yet importable.

add get_cuda_native_handle

8970dca

leofang added this to the cuda-python 12-next, 11-next milestone Jul 22, 2025

leofang requested review from rwgk, kkraus14 and oleksandr-pavlyk July 22, 2025 13:46

leofang self-assigned this Jul 22, 2025

leofang added P1 Medium priority - Should do feature New feature or request cuda.bindings Everything related to the cuda.bindings module labels Jul 22, 2025

github-project-automation bot moved this to Todo in CCCL Jul 22, 2025

github-project-automation bot added this to CCCL Jul 22, 2025

ensure the path variable is from the same scope

4ca551c

leofang commented Jul 22, 2025

View reviewed changes

This comment has been minimized.

Sign in to view

leofang requested a review from shwina July 23, 2025 01:39

rwgk previously approved these changes Jul 23, 2025

View reviewed changes

cuda_bindings/tests/test_utils.py Outdated Show resolved Hide resolved

github-project-automation bot moved this from Todo to In Review in CCCL Jul 23, 2025

address review comment + add missing API ref

c467526

leofang dismissed rwgk’s stale review via c467526 July 23, 2025 09:21

fix linter error

53055e5

rwgk approved these changes Jul 23, 2025

View reviewed changes

leofang merged commit 8cc1b68 into NVIDIA:main Jul 23, 2025
53 checks passed

leofang mentioned this pull request Jul 23, 2025

Add a deprecation warning for __int__() #775

Open

leofang deleted the utils_get_handle branch July 23, 2025 17:05

rwgk added a commit to rwgk/cuda-python that referenced this pull request Jul 24, 2025

Simplify cuda_bindings/site-packages/_cuda_bindings_redirector.py (as…

7d53118

… suggested by ChatGPT) to resolve NVIDIA#773 (comment)

rwgk mentioned this pull request Jul 24, 2025

_cuda_bindings_redirector.py fix (made necessary by PR #773) #780

Merged

Add get_cuda_native_handle #773

Add get_cuda_native_handle #773

Uh oh!

Conversation

leofang commented Jul 22, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

copy-pr-bot bot commented Jul 22, 2025

Uh oh!

leofang commented Jul 22, 2025

Uh oh!

leofang commented Jul 22, 2025

Uh oh!

leofang commented Jul 22, 2025

Uh oh!

leofang Jul 22, 2025

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

leofang commented Jul 23, 2025

Uh oh!

rwgk left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

leofang commented Jul 23, 2025

Uh oh!

Uh oh!

rwgk commented Jul 23, 2025

Uh oh!

github-actions bot commented Jul 23, 2025

Uh oh!

leofang commented Jul 23, 2025

Uh oh!

rwgk commented Jul 23, 2025

Uh oh!

leofang commented Jul 23, 2025

Uh oh!

rwgk commented Jul 23, 2025

Uh oh!

Uh oh!

Add `get_cuda_native_handle` #773

Add `get_cuda_native_handle` #773

leofang commented Jul 22, 2025 •

edited

Loading