Update `cuda.bindings` to 13.0.0 #792

leofang · 2025-08-04T15:46:03Z

Description

Close #791.

Checklist

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

Rebase to CTK 13.0 RC14

…ersion_13 Make previously overlooked 12.9.0 → 13.0.0 changes

* Update SUPPORTED_WINDOWS_DLLS: kitpicks/cuda-r13-0/13.0.0/013/local_installers/cuda_13.0.0_windows.exe * Update SUPPORTED_LINUX_SONAMES: kitpicks/cuda-r13-0/13.0.0/013/local_installers/cuda_13.0.0_580.31_linux.run * 013 → 014: SUPPORTED_LINUX_SONAMES unchanged * 013 → 014: SUPPORTED_WINDOWS_DLLS unchanged * cybind update with 13.0.0 headers (014) * Bump cuda/bindings/_version.py → 13.0.0 * test_nvjitlink.py: remove sm_60, add sm_100 * Updates from cybind after removing all 11.x headers (affects "automatically generated" comments only). * Add new toolshed/reformat_cuda_enums_as_py.py (reads cuda.h, driver_types.h headers directly). * Use new toolshed/reformat_cuda_enums_as_py.py to regenerate driver_cu_result_explanations.py, runtime_cuda_error_explanations.py * Use `driver.cuDeviceGetUuid()` instead of `driver.cuDeviceGetUuid_v2()` with CTK 13. * Adjustments for locating nvvm directory in CTK 13 installations.

* Add missing error handling (tests/test_nvjitlink.py) * Add missing `const` in cudaMemcpyBatchAsync call (cuda/bindings/runtime.pyx.in) * Add qa/13.0.0/01_linux.sh * Remove qa/13.0.0/01_linux.sh after it was moved to a new upstream qa branch. * Strictly correct casts for cudaMemcpyBatchAsync (generated by cython_gen). * Pragmatic minimal fix for cudaMemcpyBatchAsync casts (works with Linux and Windows). (generated with cython-gen)

…am bindings" This reverts commit 104abbd.

Fix accident from updating `SUPPORTED_WINDOWS_DLLS` for CTK 13

Cuda 13 soname updates

) * Linux update from cuda_13.0.0_580.46_kitpicks025_linux.run: no-op b/o NVIDIA/cuda-python-private#95 * Windows update from cuda_13.0.0_kitpicks025_windows.exe

…s overlooked. Direct commit for simplicity.

…VIDIA#94) * CCCL_INCLUDE_PATH fixes in test_event.py, test_launcher.py * Add new file (accidentally missing in a prior commit). * Fix pre-commit errors in new tests/helpers.py * 12→13 compatibility fixes in cuda/core/experimental/_graph.py * CTK 12 compatibility (tests/test_cuda_utils.py) * Make the cuda/core/experimental/_graph.py changes backwards compatible. * Do not try to hide `13` in cuda_core/tests/test_cuda_utils.py * More elegant handling of `CCCL_INCLUDE_PATHS` in cuda_core/tests/helpers.py * Remove stray empty line (cuda_core/tests/conftest.py). * Fix logic error computing CCCL_INCLUDE_PATHS in cuda_core/tests/helpers.py

rwgk · 2025-08-06T02:35:28Z

/ok to test

copy-pr-bot · 2025-08-06T02:35:32Z

/ok to test

@rwgk, there was an error processing your request: E1

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/

rwgk · 2025-08-06T02:36:01Z

/ok to test de1e83a

rwgk · 2025-08-06T02:59:46Z

There is still a problem with 13.0.0 wheels, both Linux and Windows, but I think for nvvm only.

I'm puzzled, because it works interactively (I tried both Linux and Windows). I'll continue working on this in about one hour; good chance that I can keep working on it then until I wrestle this down for good.

rwgk · 2025-08-06T04:52:46Z

/ok to test e18a5e8

rwgk · 2025-08-06T04:58:07Z

I started the testing to see where we stand with the CI (mainly with Windows).

I know from local testing that we're still up against this error (Linux):

tests/test_kernelParams.py::test_kernelParams_empty nvrtc: error: failed to open libnvrtc-builtins.so.13.0.
  Make sure that libnvrtc-builtins.so.13.0 is installed correctly.^@
FAILED

I need to work on understanding why this happens, because this looks as expected:

rwgk-win11.localdomain:~/forked/cuda-python $ ll Cp13WslVenv/lib/python3.12/site-packages/nvidia/cu13/lib/
total 375M
-rw-r--r-- 1 rgrossekunst rgrossekunst 3.1M Aug  5 21:39 libcufile.so.0
-rw-r--r-- 1 rgrossekunst rgrossekunst  43K Aug  5 21:39 libcufile_rdma.so.1
-rw-r--r-- 1 rgrossekunst rgrossekunst  95M Aug  5 21:39 libnvJitLink.so.13
-rw-r--r-- 1 rgrossekunst rgrossekunst 4.2M Aug  5 21:39 libnvrtc-builtins.alt.so.13.0
-rw-r--r-- 1 rgrossekunst rgrossekunst 4.2M Aug  5 21:39 libnvrtc-builtins.so.13.0
-rw-r--r-- 1 rgrossekunst rgrossekunst 105M Aug  5 21:39 libnvrtc.alt.so.13
-rw-r--r-- 1 rgrossekunst rgrossekunst 105M Aug  5 21:39 libnvrtc.so.13
-rw-r--r-- 1 rgrossekunst rgrossekunst  61M Aug  5 21:42 libnvvm.so.4
rwgk-win11.localdomain:~/forked/cuda-python $

But why then the error?

rwgk · 2025-08-06T05:28:45Z

Windows passes!

But for Linux, I think there is a bug/oversight in the nvidia_cuda_nvrtc wheel.

For comparison, with CTK 12:

(Ctk12NvidiaWheelsVenv) rwgk-win11.localdomain:~ $ pip install "nvidia-cuda-nvrtc-cu12"
Collecting nvidia-cuda-nvrtc-cu12
  Using cached nvidia_cuda_nvrtc_cu12-12.9.86-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl.metadata (1.7 kB)
Using cached nvidia_cuda_nvrtc_cu12-12.9.86-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl (89.6 MB)
Installing collected packages: nvidia-cuda-nvrtc-cu12
Successfully installed nvidia-cuda-nvrtc-cu12-12.9.86

(Ctk12NvidiaWheelsVenv) rwgk-win11.localdomain:~/Ctk12NvidiaWheelsVenv/lib/python3.12/site-packages/nvidia/cuda_nvrtc/lib $ readelf -d libnvrtc.so.12 | grep -E 'RPATH|RUNPATH'
 0x000000000000001d (RUNPATH)            Library runpath: [$ORIGIN:]
(Ctk12NvidiaWheelsVenv) rwgk-win11.localdomain:~/Ctk12NvidiaWheelsVenv/lib/python3.12/site-packages/nvidia/cuda_nvrtc/lib

Now with CTK 13:

(Ctk13NvidiaWheelsVenv) rwgk-win11.localdomain:~ $ pip install "nvidia-cuda-nvrtc~=13.0"
Collecting nvidia-cuda-nvrtc~=13.0
  Using cached nvidia_cuda_nvrtc-13.0.48-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl.metadata (1.7 kB)
Using cached nvidia_cuda_nvrtc-13.0.48-py3-none-manylinux2010_x86_64.manylinux_2_12_x86_64.whl (90.2 MB)
Installing collected packages: nvidia-cuda-nvrtc
Successfully installed nvidia-cuda-nvrtc-13.0.48

(Ctk13NvidiaWheelsVenv) rwgk-win11.localdomain:~/Ctk13NvidiaWheelsVenv/lib/python3.12/site-packages/nvidia/cu13/lib $ readelf -d libnvrtc.so.13 |
 grep -E 'RPATH|RUNPATH'
(Ctk13NvidiaWheelsVenv) rwgk-win11.localdomain:~/Ctk13NvidiaWheelsVenv/lib/python3.12/site-packages/nvidia/cu13/lib $

I'm afraid we have to work around that somehow.

To resolve this issue: NVIDIA#792 (comment)

rwgk · 2025-08-06T06:56:10Z

/ok to test ff339b6

rwgk · 2025-08-06T07:00:28Z

For completeness, this ChatGPT conversation explains how I arrived at commit ff339b6:

https://chatgpt.com/share/6892fd31-aafc-8008-a461-58ff0c602a0a

…refresh (no-op)

…no-op)

kkraus14 · 2025-08-06T13:23:16Z

/ok to test 37ef8e0

leofang · 2025-08-06T15:27:29Z

Since the CI was already green and the latest changes were doc-only, let me admin-merge this and run the CI in #795.

github-actions · 2025-08-06T15:41:52Z

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

vzhurba01 and others added 30 commits May 27, 2025 14:48

Propagate generated path finder changes

78b89cb

Rebase to 13.0 RC14

b423deb

Update license headers on examples

6413bab

Regenerate after merging upstream

1773388

Update benchmarks and run pre-commit

af48e18

Merge pull request NVIDIA#79 from vzhurba01/69-bindings-ctk-13.0

2dbc870

Rebase to CTK 13.0 RC14

cython-gen output, NO manual changes

6509f3c

Bump cuda/bindings/_version.py → 13.0.0

3d33553

Merge pull request NVIDIA#82 from rwgk/automatically_generated_with_v…

1d45862

…ersion_13 Make previously overlooked 12.9.0 → 13.0.0 changes

Merge branch 'world-main' into unreleased-13.0

6f1eacf

Merge branch 'world-main' into unreleased-13.0

033512b

Merge branch 'world-main' into unreleased-13.0

47622f3

Merge branch 'world-main' into unreleased-13.0

0aa9a1f

Merge branch 'world-main' into unreleased-13.0

2ac7571

Merge branch 'world-main' into unreleased-13.0

4eb228a

print *prog pointers in nvrtcCreateProgram, nvrtcCompileProgram bindings

104abbd

Remove stray "nvrtc64_*_0.alt.dll" entries in SUPPORTED_WINDOWS_DLLS

f64edeb

Revert "print *prog pointers in nvrtcCreateProgram, nvrtcCompileProgr…

50e41bb

…am bindings" This reverts commit 104abbd.

Merge pull request NVIDIA#92 from rwgk/unreleased-13.0_dbg01

de74968

Fix accident from updating `SUPPORTED_WINDOWS_DLLS` for CTK 13

Merge branch 'world-main' into unreleased-13.0

d8ff357

_find_lib_dir_using_cuda_home(): Windows CTK 13 → bin\x64

eaf66f2

getLocalRuntimeVersion(): Search for libcudart.so.13

ebc9920

SUPPORTED_LINUX_SONAMES: Add CTK 13 soname values

d760969

Merge pull request NVIDIA#95 from robertmaynard/cuda-13-soname-updates

d67a293

Cuda 13 soname updates

Update path_finder/supported_libs.py from kitpicks 13.0.0/025 (NVIDIA#96

7f8fa90

) * Linux update from cuda_13.0.0_580.46_kitpicks025_linux.run: no-op b/o NVIDIA/cuda-python-private#95 * Windows update from cuda_13.0.0_kitpicks025_windows.exe

Merge branch 'world-main' into unreleased-13.0

718b3db

This trivial change should have been included in PR NVIDIA#81, but wa…

3bec7e3

…s overlooked. Direct commit for simplicity.

Fix Windows site-packages search for CTK 13

de1e83a

Also add "nvidia-nvvm~=13.0" in cuda_bindings/pyproject.toml

e18a5e8

Add _work_around_known_bugs() in load_dl_linux.py

ff339b6

To resolve this issue: NVIDIA#792 (comment)

rwgk added 4 commits August 6, 2025 01:59

driver_cu_result_explanations.py, runtime_cuda_error_explanations.py …

25d4b0a

…refresh (no-op)

SUPPORTED_LINUX_SONAMES refresh (no-op)

db83fbe

SUPPORTED_WINDOWS_DLLS refresh (no-op)

36d9f06

Update generated files: nvjitlink, nvvm (trivial changes, functional …

37ef8e0

…no-op)

kkraus14 previously approved these changes Aug 6, 2025

View reviewed changes

github-project-automation bot moved this from Todo to In Review in CCCL Aug 6, 2025

leofang marked this pull request as ready for review August 6, 2025 15:04

update release dates

ba95fc9

leofang dismissed kkraus14’s stale review via ba95fc9 August 6, 2025 15:18

kkraus14 approved these changes Aug 6, 2025

View reviewed changes

leofang merged commit c016d65 into NVIDIA:main Aug 6, 2025
1 check passed

github-project-automation bot moved this from In Review to Done in CCCL Aug 6, 2025

leofang deleted the unreleased-13.0 branch August 6, 2025 15:27

leofang linked an issue Aug 6, 2025 that may be closed by this pull request

[BUG]: cuda_bindings/examples globalToShmemAsyncCopy_test.py "catastrophic error" is masked by pytest skipped #730

Closed

1 task

leofang mentioned this pull request Aug 6, 2025

[BUG]: cuda_bindings/examples globalToShmemAsyncCopy_test.py "catastrophic error" is masked by pytest skipped #730

Closed

1 task

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Update `cuda.bindings` to 13.0.0 #792

Update `cuda.bindings` to 13.0.0 #792

Uh oh!

leofang commented Aug 4, 2025 •

edited

Loading

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

copy-pr-bot bot commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025 •

edited

Loading

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

kkraus14 commented Aug 6, 2025

Uh oh!

leofang commented Aug 6, 2025

Uh oh!

Uh oh!

github-actions bot commented Aug 6, 2025

Uh oh!

Uh oh!

Update cuda.bindings to 13.0.0 #792

Update cuda.bindings to 13.0.0 #792

Uh oh!

Conversation

leofang commented Aug 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

copy-pr-bot bot commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

rwgk commented Aug 6, 2025

Uh oh!

kkraus14 commented Aug 6, 2025

Uh oh!

leofang commented Aug 6, 2025

Uh oh!

Uh oh!

github-actions bot commented Aug 6, 2025

Uh oh!

Uh oh!

Update `cuda.bindings` to 13.0.0 #792

Update `cuda.bindings` to 13.0.0 #792

leofang commented Aug 4, 2025 •

edited

Loading

rwgk commented Aug 6, 2025 •

edited

Loading