Establish path finder support policy

[leofang]

Capturing some thoughts discussed offline

• Library teams can request support for their libraries (shipped outside of CTK), provided that they contribute and help maintain the search logics.

For non-CTK libraries we need each library team’s help if they need pathfinder’s support. We will guide them for onboarding, review their contributions, and make sure we cut a new release to meet their release schedule (we’re very agile in this). We also need the library teams to be responsible for future maintenance.

• ...

[ZzEeKkAa]

One of the use cases: #776 (libmathdx)

[ZzEeKkAa]

cudss might be another use case: #833

[ZzEeKkAa]

cufftmp is another use case https://docs.nvidia.com/cuda/cufftmp/index.html

[rwgk]

cufftmp is another use case https://docs.nvidia.com/cuda/cufftmp/index.html

@ZzEeKkAa How is that library usually installed?

(I clicked around for a few minutes, I found examples, but didn't get lucky finding instructions for installing a .so or .dll.)

[ZzEeKkAa]

cufftmp is another use case https://docs.nvidia.com/cuda/cufftmp/index.html

@ZzEeKkAa How is that library usually installed?

(I clicked around for a few minutes, I found examples, but didn't get lucky finding instructions for installing a .so or .dll.)

It is only for linux, if I understand correctly.

You can install it by extracting tarball from https://developer.nvidia.com/cufftmp-downloads and populating LD_LIBRARY_PATH. Other option is to use nvidia-cufftmp-cu12 wheel

[rwgk]

You can install it by extracting tarball from https://developer.nvidia.com/cufftmp-downloads and populating LD_LIBRARY_PATH.

I found a .deb there and installed it following the instructions exactly. Then, using PR #834, this works even without populating LD_LIBRARY_PATH:

(Cp13WslVenv) rwgk-win11.localdomain:~/forked/cuda-python $ python
Python 3.12.3 (main, Jun 18 2025, 17:59:45) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> from cuda.pathfinder import load_nvidia_dynamic_lib
>>> load_nvidia_dynamic_lib("cufftMp")
LoadedDL(abs_path='/lib/x86_64-linux-gnu/libcufftMp.so', was_already_loaded_from_elsewhere=False, _handle_uint=694625936)
>>>

It works through the standard system search.

There is a softlink:libcufftMp.so → libcufftMp.so.11

The details of the softlink involve /etc/alternatives, I guess that's to support switching versions.