Adds path for the CUDA static library based on CUDA_HOME #837
Conversation
|
@Andy-Jost to trigger CI to run you need to comment "/ok to test". More info here: https://docs.gha-runners.nvidia.com/platform/apps/copy-pr-bot/faqs |
|
/ok to test 6b178c1 |
|
This breaks Windows builds because LIBRARY_PATH is specific to Linux. While I could place a guard around the check, it's probably better to close this PR without merging. |
I think it'd be lovely to complete this PR, I stumbled over this many times. Under Windows, we need This is a good way to test for Windows (we use that in |
I am looking into a solution that works for Windows as well as Linux. |
|
Probably just update |
leofang
added
enhancement
Andy-Jost
force-pushed
the
fix/add-library-path-check
branch
from
37e877d to
44579bc
Compare
|
The latest changes eliminate the need to set either LIBRARY_PATH or LIB and should resolve issue #608. Where can I find the source for the installation instructions? I.e., https://nvidia.github.io/cuda-python/cuda-bindings/latest/install.html. Those can be updated now. |
This did the trick. |
Andy-Jost
force-pushed
the
fix/add-library-path-check
branch
2 times, most recently
from
d62461d to
96f3514
Compare
|
/ok to test 96f3514 |
This comment has been minimized.
This comment has been minimized.
Andy-Jost
force-pushed
the
fix/add-library-path-check
branch
from
96f3514 to
49ec36d
Compare
Andy-Jost
changed the title
|
Thanks, @Andy-Jost! The diff is clean now. I think to test this properly in the CI, we should delete these two lines cuda-python/.github/workflows/build-wheel.yml Line 160 in b400e34 cuda-python/.github/workflows/build-wheel.yml Line 164 in b400e34 Not sure if I miss any. I hope not 🤔 |
cuda_bindings/setup.py
Outdated
| os.path.dirname(sysconfig.get_path("include")), | ||
| ] + include_path_list | ||
| library_dirs = [sysconfig.get_path("platlib"), os.path.join(os.sys.prefix, "lib")] | ||
| cudalib_subdir = r"lib\x64" if sys.platform == "win32" else "lib64" |
There was a problem hiding this comment.
On the Linux side, one annoying thing is that depending on how CUDA is installed, this could be lib or lib64:
$ ls -l /usr/local/cuda-13.0/lib64/libcudart_static.a
-rw-r--r-- 1 root root 1361104 Aug 5 09:32 /usr/local/cuda-13.0/lib64/libcudart_static.a
$ ls -l /usr/local/cuda-13.0/targets/x86_64-linux/lib/libcudart_static.a
-rw-r--r-- 1 root root 1361104 Aug 5 09:32 /usr/local/cuda-13.0/targets/x86_64-linux/lib/libcudart_static.aWe probably should support both.
|
/ok to test 19aaca9 |
|
I see a failure in pre-commit.ci due to "ruff format" creating diffs. I think I get the general idea, but I don't see how to get more information (e.g., see the diffs, or run the ruff format locally). I can probably fix the formatting issue by guessing, but in general how can I get the details? |
|
/ok to test 19aaca9 |
|
/ok to test 26da853 |
Andy-Jost
force-pushed
the
fix/add-library-path-check
branch
from
26da853 to
73acddf
Compare
|
/ok to test 73acddf |
|
|
@copilot please backport this PR to the 12.9.x branch. (Since this PR touched some files in the .github folder, the backport bot does not work.) |
leofang
added
to-be-backported
|
Git push to origin failed for 12.9.x with exitcode 1 |
|
Git push to origin failed for 12.9.x with exitcode 1 |
Co-authored-by: leofang <5534781+leofang@users.noreply.github.com>
…bc-85a8c5353e08 Backport PR #837: Add path for CUDA static library based on CUDA_HOME
4 participants
Both CUDA_HOME (or CUDA_PATH) and LIBRARY_PATH are required to be set in the environment during installation. There was already a check for CUDA_HOME in setup.py, but no check for LIBRARY_PATH. This change adds a check to make sure LIBRARY_PATH is set.
When LIBRARY_PATH is not set, the installation fails at link time and indicates that cudart_static could not be found.