On 25 September 2025 at 22:51, Christian Kastner wrote: | On 2025-09-25 21:47, Dirk Eddelbuettel wrote: | > Hm. I do not think that is very clever. It basically just renders the library | > much less deployable. | > | > And I am not being argumentative here but why did you bother packaging it | > (which from the complexity of the upstream setup is surely non-trivial work) | > when at the end of the day you do not want user of the library use it? | | The goal was to ship the llama.cpp *utilities*. In the first version of | this [1], there was only bin:llama.cpp, with no library or -dev packages | at all.
Ack. And that makes sense. | ggml was factored out because it is the most complex part of the package | (building various backends and so on), and it is also needed by | whisper.cpp. We didn't want to duplicate the work. And this is still | experimental; it's still possible that we might need to embed again | after all. | | The library and -dev packages were split out by request because one big | monolith package with every possible component and test is sub-optimal | to someone who just needs llama-server, for example. This request seemed | reasonable. | | And it's prep work for the time when the libraries actually become | stable. All good. | > I am being serious here. I was planning to work "on top" and now I can't | > because I would have to re-invent library discovery on every possible distro | > or deployment. | | Lacking stability, you probably couldn't do that anyway, even if every | distro shared the same installation layout. You'd still have to work out | how to deal with different versions. | | I understand that you have a use case for the libraries, but I don't | understand the challenge here. Whatever you wanted to do, you | can still do, just add one more argument to pkg-config or cmake. Without the pkgconf redirection, how would I know where some other system or OS keeps the libraries? | > I guess I could hardcode those paths | | Those private paths won't change, if that helps. Ok, and for now 'deployment on Debian/Ubuntu' is like a requirement for my (as of yet non-existing) experiment anyway. | > In reality the library likely moves too fast anyway. | | Yes, that is my point. I (or any other deployment) can't give you | predictability/stability if it's not there to give in the first place. That's ok. My risk in looking foolish in they eyes of any downstream users/ | Best, | Christian | | PS: I'll be updating llama.cpp and ggml to recent versions over the weekend. Much appreciated. My dirty secret is that I am currently on Ubuntu 25.04 anyway which doesn't have it yet so this may end up Docker / devcontainer based anyway. Again, really appreciate that you packaged this. Cheers, Dirk -- dirk.eddelbuettel.com | @eddelbuettel | [email protected]
