Add multi-cuda wheel build #289
Conversation
|
/ok to test 1ab0a2e |
shwina
force-pushed
the
multi-cuda-wheel
branch
from
1ab0a2e to
4b5fb32
Compare
|
/ok to test 4b5fb32 |
|
/ok to test a020600 |
|
/ok to test 18ecf11 |
|
/ok to test b62956b |
|
/ok to test edd31c7 |
|
/ok to test d178a08 |
|
/ok to test fd2de8c |
This reverts commit edd31c7.
|
/ok to test 1e2ef0f |
|
/ok to test d54f264 |
|
/ok to test 13cb606 |
|
/ok to test 9af979d |
This reverts commit 13cb606.
|
/ok to test ec3615b |
|
|
||
| [project.optional-dependencies] | ||
| # CUDA 12.x dependencies | ||
| cu12 = ["cuda-bindings>=12.0.0,<13.0.0", "nvidia-cuda-cupti-cu12"] |
There was a problem hiding this comment.
Remove the cuda-bindings dependency as it's just used as a proxy for the CUDA major version right now.
alliepiper
left a comment
alliepiper
left a comment
There was a problem hiding this comment.
CMake / infra side look good to me
| CPMAddPackage("gh:pybind/pybind11@3.0.0") | ||
|
|
||
| # Determine CUDA major version for directory structure | ||
| string(REGEX MATCH "^([0-9]+)" CUDA_VERSION_MAJOR "${CUDAToolkit_VERSION}") |
There was a problem hiding this comment.
nit: CUDAToolkit_VERSION_MAJOR should already be defined as part of CUDAToolkit.
| if [[ "$(uname -m)" == "aarch64" ]]; then | ||
| readonly cuda12_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda12_version}-${devcontainer_distro}-py${py_version}-arm64 | ||
| readonly cuda13_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda13_version}-${devcontainer_distro}-py${py_version}-arm64 | ||
| else | ||
| readonly cuda12_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda12_version}-${devcontainer_distro}-py${py_version} | ||
| readonly cuda13_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda13_version}-${devcontainer_distro}-py${py_version} | ||
| fi |
There was a problem hiding this comment.
Nit: Perhaps introduce readonly host_arch_suffix to set to -arm64 or leave empty` and then image names could be set unconditionally, and arch names would only need to be changed in one place:
readonly cuda12_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda12_version}-${devcontainer_distro}-py${py_version}${arch_suffix}
readonly cuda13_image=rapidsai/ci-wheel:${devcontainer_version}-cuda${cuda13_version}-${devcontainer_distro}-py${py_version}${arch_suffix}| # Test dependencies for CUDA 13 | ||
| test-cu13 = ["pynvbench[cu13]", "pytest", "cupy-cuda13x", "numba"] | ||
|
|
||
| # Generic test dependencies (defaults to CUDA 12) |
There was a problem hiding this comment.
Shouldn't we default to the latest major version of CTK? Is there a reason not to default to using cu13?
3 participants
This PR introduces "multi-CUDA" wheels builds, exactly similar to the ones we build for CCCL, as described here.
Instead of shipping multiple packages (
pynvbench-cu12,pynvbench-cu13); we ship a single one (pynvbench) that contains both CUDA 12 and 13 binaries (bindings). The resulting wheel size is just ~2.2MB, so this is not a problem. At runtime, depending on the available CUDA version, the appropriate binary is loaded.At install time, we still need the user to specify which CUDA major version to install for - because the version of required dependencies (
nvidia-cuda-cupti-cu12) depend on the CUDA major version. This specification of the CUDA version is done using "extras":