Dask-CUDA: CUDA 12 Conda Packages

[jakirkham]

Pushing for CUDA 12 Conda package builds of libraries for RAPIDS 23.04. This issue is for tracking that work.

Requirements (please recheck):

Developer tasks

Beta Give feedback

1. I've confirmed the list below is correct (or updated if not)

Dependencies from Conda

Beta Give feedback

1. pynvml v11.5.0 conda-forge/pynvml-feedstock#11
   +0
2. Numba: Support CUDA 12 conda-forge/numba-feedstock#112
   +0

RAPIDS Conda dependencies

Beta Give feedback

1. RMM: CUDA 12 Conda Packages rmm#1207
   6 of 6
   +0
   
2. cuDF: CUDA 12 Conda Packages cudf#12771
   15 of 15
   +0
   
3. KvikIO: Build CUDA 12 packages kvikio#224
   feature request non-breaking +2
   
4. UCX-Py: CUDA 12 Conda Packages ucx-py#927
   7 of 7
   +0
   

Optional dependencies from Conda

Beta Give feedback

1. CuPy: Support CUDA 12 conda-forge/cupy-feedstock#139
   +0
2. nvtx (Python)
3. UCX: Support CUDA 12 conda-forge/ucx-split-feedstock#116
   +0

[wence-]

(via distributed/ucx-py)

Not sure what this is supposed to mean. Do you just want to write ucx-py (RAPIDS) ?

Also not sure how you're specifying transitive dependencies, e.g. if you're optionally depending on cudf, do you also need to depend on rmm (which is a transitive dep of cudf)?

[jakirkham]

^ @pentschev, do you have thoughts on this question?

[pentschev]

(via distributed/ucx-py)

Not sure what this is supposed to mean. Do you just want to write ucx-py (RAPIDS) ?

I just tried to be somewhat verbose, not even sure this is a necessity. What I meant to say is that the dependency is Dask-CUDA->Distributed->UCX-Py->UCX, and it is an important dependency for us, thus the verbosity. I'll defer to @jakirkham on how that can be best viewed.

Also not sure how you're specifying transitive dependencies, e.g. if you're optionally depending on cudf, do you also need to depend on rmm (which is a transitive dep of cudf)?

Again just for verbosity. Users may have cuDF workflows that depend on RMM, or just CuPy (for example) workflows that also depend on RMM. With that said, I just wanted to make sure this is properly visualized, if this is only overhead I'm fine reorganizing or removing transitive dependencies altogether, probably @jakirkham has a best idea of what's important and could remove any overheads.

[jakirkham]

Yeah no strong feelings here. Understand what is written above. Ok with the other alternatives mentioned in the thread

[jakirkham]

In PR ( #1130 ), we are trying to update to PyNVML 11.5.0, but are having some issues ( #1130 (comment) )

AIUI PyNVML 11.5.0 is compatible with CUDA 12. Checking to see if an older PyNVML version would suffice

[jakirkham]

After checking offline, have confirmed we need PyNVML 11.5.0+ for CUDA 12 support.