added kokkos to libraries.yaml in the cuda section

[rbourgeois33]

See the corresponding PR

[mattgodbolt]

I'm trying to understand what this is doing: it seems to duplicate the kokkos we already have elsewhere? They both install to the same directory so... the second copy does nothing:

 $ ce_install install kokkos 4.6.01
2025-06-16 09:45:37,614 lib.installation_context INFO     Making uncached requests
2025-06-16 09:45:37,615 lib.ce_install  INFO     Creating thread pool with 8 workers
Installing libraries/c++/kokkos 4.6.01
2025-06-16 09:45:43,958 lib.ce_install  INFO     libraries/c++/kokkos 4.6.01 is already installed, skipping
Installing libraries/kokkos 4.6.01
2025-06-16 09:45:43,958 lib.ce_install  INFO     libraries/kokkos 4.6.01 is already installed, skipping
0 packages installed OK, 2 skipped, and 0 failed installation
bollo:~/d/c/infra (rbourgeois33-main|✔) 6.9s $ 

Can you explain what you're trying to achieve here? Do we need to change our existing kokkos to build with extra flags? Do we really need two separate copies of the Kokkos library? If so we'll need to change some paths.

Or am I missing something? :)

[rbourgeois33]

As discussed in the discord:

Kokkos is a compile-time polymorphic library. It requires separate build for separate backends (one for CPU, one for GPU)

We can keep one "clone" of kokkos but we do need two separate builds/installs.

Practically, this can be done in seperate build folders with e.g. cmake -B build_gpu

If a new GPU architecture needs to be supported, a third install will be necessary (and so forth)

[dalg24]

Can you post a link to the discord discussion? If the command to install a C++ library VS CUDA C++ library does not specify the language then I see how that will be ambiguous and we probably need to suffix the new package (kokkos -> kokkos-cuda)

[rbourgeois33]

Here it is:

https://discord.gg/B5WacA7

channel: ce-dev-discussions

[mattgodbolt]

I'm sorry I didn't get back to this! I've been at a conference and I'll be on holiday next week. Will try and get this working..

[rbourgeois33]

I'm sorry I didn't get back to this! I've been at a conference and I'll be on holiday next week. Will try and get this working..

No rush ! Thank again for taking time to do this. The Kokkos community is thankful :)

No guarantee it'll work exactly as intended as I doubt our builder box has the GPU stuff installed, but this will stop it trying to install over the top of the "normal" kokkos.

Let me know if my last commit does the job, maybe I should have indented the following lines ?

Cheers !

[mattgodbolt]

Thanks @rbourgeois33 that does the trick. Seems we still need some infrastrucutre changes on our side too though

[tpadioleau]

Thanks @rbourgeois33 that does the trick. Seems we still need some infrastrucutre changes on our side too though

Hi @mattgodbolt! Do you have an update on the status of those changes ?

[mattgodbolt]

I'm sorry, no, it's not been something we've prioritised! I appreciate that's frustrating, and it being infrastructurla makes it a little harder to solicit external help (though the code is OSS if you want to poke around; likely as not @partouf or I need to have a think about it though)

[tpadioleau]

I'm sorry, no, it's not been something we've prioritised! I appreciate that's frustrating, and it being infrastructurla makes it a little harder to solicit external help (though the code is OSS if you want to poke around; likely as not @partouf or I need to have a think about it though)

Don't worry we understand. We are wondering if there could be any trick on the library side that could ease the deployment. For example if we rename our library, let's say kokkos-cuda, could it be helpful ?