[FEA]: Constraints between library versions

[jmulcahy]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

cuda.bindings

Is your feature request related to a problem? Please describe.

I would like some method to ensure that all CUDA libraries have matching major and minor versions in order to ensure compatibility.

For example, if I (on Linux) pip install torch "numba_cuda[cu12]" nvidia-cuda-nvcc-cu12 I will get a mixture of CUDA 12.6 packages from torch and nvidia-cuda-nvcc-cu12 12.9. The version mismatch breaks numba_cuda's compilation.

Describe the solution you'd like

I would like major and minor versions of CUDA packages to be able to be constrained together.

This might look like pip install torch "numba_cuda[cu12]" "cuda-bindings[all]" resolving nvidia-cuda-nvcc-cu12 to 12.6.*.

Describe alternatives you've considered

I've tried a few different packages that I hoped to apply this constraint somehow. None seemed to. Part of the complication here is that PyTorch manages CUDA versions with an index flag that can't be embedded in pyproject.toml without the use of other tools.

Additional context

This could possibly be a numba_cuda issue, but it's not clear what that package could do about this situation. For example, numba_cuda[cu12] tries to constrain to CUDA 12.9 by setting cuda-bindings==12.9.* in its dependencies. However, pip install torch "numba_cuda[cu12]" installs nvidia-cuda-nvrtc-cu12 12.6 and nvidia-cuda-nvcc-cu12 12.9.

[leofang]

Hi @jmulcahy sorry for late reply. @emmatyping-nv has been working on meeting this exact needs for some time, and I am happy to share that the new cuda-toolkit metapackage is released on PyPI. It supports CUDA 11.8, 12.x, and 13.x. For example, you can build a consistent 12.9 environment by doing pip install cuda-toolkit[nvcc,nvrtc]=12.9. Please let us know if you encounter any issue. CUDA Python projects will start to use this new machinery going forward (#853).

[jmulcahy]

This is pretty nice. There does seem to be some light weirdness, although maybe that irons out as packages update dependencies.

As an example, this installs an env where numba_cuda works (on Linux where CUDA is implied for torch):

pip install torch "cuda-toolkit[nvrtc,nvcc]==12.*" "numba_cuda[cu12]"

While this installs a mishmash of CUDA 12 and 13 that doesn't work:

pip install torch "cuda-toolkit[nvrtc,nvcc]" "numba_cuda[cu12]"