Migrate samples - #2959
Open
danielfrg wants to merge 46 commits into
Open
Migrate samples#2959danielfrg wants to merge 46 commits into
danielfrg wants to merge 46 commits into
Conversation
Resolves conflicts caused by concurrent changes on main: - Relicense (NVIDIA#2293) updated SPDX headers on cuda_bindings/examples/*.py and cuda_bindings/cuda/bindings/_example_helpers/*. All of those files were removed on this branch, so main's edits are discarded via git rm. - Bindings-side test_examples.py, cuda_core/examples/pytorch_example.py and cuda_core/tests/example_tests/test_basic_examples.py were removed on this branch; main's edits to them are discarded. - scripts/run_tests.sh was removed on main (NVIDIA#2271); accept the deletion. - ci/tools/run-tests: keep this branch's optional sample-deps install (nvtx, pillow) but pick up main's rename CUDA_VER_MINOR -> TEST_CUDA_MAJOR_MINOR. - Four new texture / GL-interop examples added to cuda_core/examples/ on main (gl_interop_fluid.py, gl_interop_fluid_numba_cuda_mlir.py, gl_interop_mipmap_lod.py, texture_sample.py) are migrated into their own directories under samples/ with README.md and requirements.txt. Added a DISPLAY guard to glInteropFluid and glInteropMipmapLod so they waive with exit code 2 on headless runners, matching glInteropPlasma. - ruff.toml: ignore RUF059 under samples/** (matches the pre-existing ignore for examples/**). - .spdx-ignore: drop the obsolete cuda_bindings/examples/* line. Verified end-to-end: - pixi install -e samples: solves cleanly. - pytest cuda_core/tests/example_tests/test_samples.py: 46 passed, 9 skipped (all legitimate: no DISPLAY, no P2P, sub-Hopper GPU, or intentionally omitted deps), 0 failed. 55 samples collected. - ruff check + format: clean. - toolshed/check_spdx.py: clean on edited files.
Relative imports fail when example_tests has no __init__.py; import run_samples as a top-level module instead.
- systemInfo: guard get_process_name() call behind CUDA_BINDINGS_NVML_IS_COMPATIBLE to avoid NameError when NVML bindings are incompatible with the driver. - vectorAdd: replace cp.random.rand() with NumPy-generated data transferred via cp.asarray() to avoid a dependency on libcurand which is not present in the cuda-13x wheel environment.
CuPy's FFT backend requires libcufft (so.11 or so.12), which is not bundled in the cuda-13x wheel environment used in CI. Probe cupy.cuda.cufft at startup and exit with the runner waiver code (CUDA_PYTHON_SAMPLE_WAIVER_EXIT_CODE=77) so the test is reported as SKIPPED rather than FAILED.
…tSignalAnalysis CuPy's FFT backend requires libcufft (shipped by nvidia-cufft-cu12 on PyPI). Adding it to the PEP 723 inline script block and requirements.txt lets the sample runner automatically waive the test when cuFFT is not installed, rather than crashing with an ImportError at runtime.
…umpyVsCupy cp.dot() requires libcublasLt (shipped by nvidia-cublas-cu12 on PyPI). Adding it to the PEP 723 inline script block and requirements.txt lets the sample runner automatically waive the test when cuBLAS is not installed.
…gram cp.random requires libcurand which is not installed in the cuda-13x wheel CI environment. The random data is only used as histogram input, so generating it on CPU via numpy and uploading with cp.asarray() is equivalent and removes the curand dependency.
… is unsupported ManagedMemoryResource with preferred_location requires concurrent managed memory access support, which is not available on all devices (e.g. RTX 4090 in the CI environment). Catch CUDA_ERROR_NOT_SUPPORTED at runtime and exit with the runner waiver code so the test is reported as SKIPPED rather than FAILED. This mirrors the existing Windows waiver already in the sample.
…cess is unavailable Per CUDA docs (cuMemPoolCreate): managed memory pools require all devices to have non-zero concurrentManagedAccess; otherwise CUDA_ERROR_NOT_SUPPORTED is returned. This attribute is 0 on Windows, WSL, and some Linux VM environments. Check device.properties.concurrent_managed_access before attempting to create the ManagedMemoryResource and exit with the runner waiver code when it is False, mirroring the existing Windows waiver already in the sample.
…cess is unavailable Managed memory pools (cuMemPoolCreate) require all devices to have non-zero concurrentManagedAccess; otherwise CUDA_ERROR_NOT_SUPPORTED is returned. This attribute is 0 on Windows, WSL, and some Linux VM environments. Move the concurrent_managed_access check to just after device info is printed and simplify the message, keeping it alongside the existing Windows waiver.
NVIDIA WSL User Guide states that only legacy CUDA IPC APIs are supported from driver R510; the newer IPC memory pool API (cuMemPoolCreate with POSIX FD handles) is not supported on WSL2. Detect WSL by checking for 'microsoft' in /proc/sys/kernel/osrelease and return False from check_ipc_support() so the sample exits cleanly with a waiver instead of crashing with CUDA_ERROR_INVALID_VALUE.
…rrent managed access is unavailable Both samples use ManagedMemoryResource which requires concurrent managed memory access (concurrentManagedAccess != 0). On WSL and Windows this attribute is 0, causing CUDA_ERROR_NOT_SUPPORTED at pool creation. Add concurrent_managed_access check after device init, consistent with the existing Windows waiver and the fix applied to blurImageUnifiedMemory.
…UDA context nvmlSystemGetProcessName returns NVML_ERROR_NOT_FOUND when the process has not yet established a CUDA context (i.e. no cudaMalloc or equivalent has been called). This is expected NVML behavior per the API docs. Guard the get_process_name() call with a try/except NotFoundError so the sample gracefully prints N/A instead of crashing.
Catch OSError (FileNotFoundError is a subclass) when pyglet tries to load opengl32.dll on headless Windows CI runners (TCC/MCDM mode or missing display driver). Exit with EXIT_WAIVED so the test is skipped rather than failing. See: NVIDIA#2767
Catch OSError (FileNotFoundError is a subclass) when pyglet tries to load opengl32.dll on headless Windows CI runners (TCC/MCDM mode or missing display driver). Exit with EXIT_WAIVED so the test is skipped rather than failing. Applies the same fix to glInteropFluid, glInteropPlasma, and glInteropFluidNumbaCudaMlir (same root cause as glInteropMipmapLod). See: NVIDIA#2767
The file already defined EXIT_WAIVED at module level; the extra definition added near the imports created a duplicate that caused test_core_sample_waivers_use_negotiated_exit_code to fail (expects exactly one assignment).
PinnedMemoryResource internally calls cuMemPoolCreate; on Windows with WDDM T4 (memory_pools_supported=False) this raises CUDA_ERROR_NOT_SUPPORTED. Check device.properties.memory_pools_supported early and exit with EXIT_WAIVED so the test is skipped rather than failing.
…unsupported PinnedMemoryResource internally calls cuMemGetMemPool; on Windows with TCC mode T4 (memory_pools_supported=False) this raises CUDA_ERROR_NOT_SUPPORTED. Check device.properties.memory_pools_supported early and exit with EXIT_WAIVED so the test is skipped rather than failing.
On Python 3.13, ctypes.windll.opengl32 wraps the underlying FileNotFoundError as AttributeError: opengl32 rather than propagating the OSError directly. Extend the except clause to (OSError, AttributeError) so the waiver fires on both Python 3.12 and 3.13 Windows CI runners.
…Harness When pytest-rerunfailures retries test_main on the same instance, self.buffer from the previous run may still be set while its backing DeviceMemoryResource has already been closed by the finally block. process.start() (spawn mode) pickles self.child_main which requires serializing self.__dict__, hitting the closed MR and raising: RuntimeError: DeviceMemoryResource has been closed when serializing dict item 'buffer' Fix by popping 'buffer' from self.__dict__ at the start of each test_main run so a stale closed-MR buffer never reaches the pickler.
…nsorMap The kernel uses <cuda/barrier> from libcudacxx (CCCL). On Windows CI with pip-installed CUDA, cuda.pathfinder cannot locate the cccl include directory because nvidia-cuda-cccl-cu12 is not installed, causing NVRTC to fail with 'cannot open source file cuda/barrier'. Add nvidia-cuda-cccl-cu12 to the PEP 723 inline dependencies and requirements.txt so the sample runner installs (or waives) it, matching the same pattern used for nvidia-cufft-cu12 in fftSignalAnalysis.
cuFile stats level is process-global state (a C library global variable inside libcufile). Running test_set_stats_level concurrently with test_stats_start_stop (which also calls set_stats_level(1)) can cause a race condition where get_stats_level() returns the value set by the other test, failing the assertion. All other cuFile stats tests (test_stats_start_stop, test_get_stats_l1/l2/l3) were already marked thread_unsafe; this commit fixes the oversight for test_set_stats_level.
SMResource.split(dry_run=True) fails with cuDevSmResourceSplit when result=NULL and smCount is non-zero. This causes the auto SM-split probe in greenContext to fail on CUDA 13.1+. Waive the sample on affected drivers until the cuda.core bug is fixed.
cuda_graphs.py and memory_ops.py were migrated from cuda_core/examples/ to samples/cuda_core/ in this PR. Update the 0.3.1 release notes to use the :sample-file: role pointing to the new locations so lychee link checks pass.
* origin/main: (128 commits) chore: refresh pixi lockfiles (NVIDIA#2956) [no-ci] Improve the Pixi lockfile freshness check (NVIDIA#2942) docs(cuda.core): add 1.2.1 to the docs version switcher (NVIDIA#2944) chore: refresh pixi lockfiles (NVIDIA#2943) ci: temporarily pause GB300 nightly run (NVIDIA#2941) [doc-only] Prepare the 13.4.3 release (NVIDIA#2936) fix(bindings): cover missing helper attrs and fix CU_JIT_WALL_TIME pointer (NVIDIA#2931) cuda.core: fix pool setup and builder teardown under stream capture (NVIDIA#2838) Fix CUDA pointer and coredump attribute types (NVIDIA#2929) Use default setuptools-scm tag parsing (NVIDIA#2922) chore: refresh pixi lockfiles (NVIDIA#2927) chore: refresh pixi lockfiles (NVIDIA#2921) feat(cuda.core): support nvJitLink incremental linking (NVIDIA#2867) tests: enumerate CUDA devices for GPU-only system checks (NVIDIA#2916) Merge 13.4.2 changes into main (NVIDIA#2913) fix(cuda.core): preserve adjacency set semantics (NVIDIA#2893) Fix primary context cleanup during Python shutdown (NVIDIA#2744) Simplify and speed up param packer (NVIDIA#2904) Fix VMM handle leaks in virtual memory paths (NVIDIA#2235) [chore] Remove obsolete sccache settings (NVIDIA#2860) ... # Conflicts: # cuda_bindings/pixi.lock # cuda_core/pixi.lock
Contributor
github-actions
Bot
added
CI/CD
danielfrg
added
P0
Revert the cp.asarray(np.random...) CPU-generate-then-transfer workaround back to cp.random.rand()/cp.random.randint() in both samples, per review feedback that the CPU fallback deviated from the original CuPy-based content and taught users an unnecessary host-to-device transfer. cp.random dlopens libcurand at runtime, which the cuda_core CI test environment does not currently install. Rather than widening the shared test-cu12/test-cu13 dependency groups, declare nvidia-curand explicitly in each sample's PEP 723 script header and requirements.txt so the shared sample runner's dependency gate (missing_dependencies) waives the sample cleanly when curand is unavailable instead of the process crashing on dlopen.
Contributor
|
/ok to test 6b5e79d |
Contributor
|
Replace nvidia-cufft-cu12/nvidia-cublas-cu12 with nvidia-cufft/nvidia-cublas in fftSignalAnalysis and numpyVsCupy, per review feedback not to mix CUDA 12 and CUDA 13 packages. Follow-up on switching to cupy-cuda13x[ctk] tracked in NVIDIA#2961.
Contributor
|
/ok to test 562e4c6 |
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
close #2866
Moving the cuda-python samples from the cuda-samples repo.
Migrating also the sample runners and added a small pytest wrapper to run the test.
Let me know if we want to do any changes to when these are run.
Checklist