Benjy, commenting on your request for review. **kernel-level execution** We want our Linux kernel to be highly deterministic, use minimal RAM and compute, have source code verifiable with real eyeballs, and be sourced from kernel.org. Process, memory, and I/O management are the Linux tour de force and it will be hard to justify using a Dual AMD GPU for arguably marginal performance improvement. If instead you are exposing a user-space LLM API through a GPU driver, then you have a case for a kernel module.
**submitting for review** That's not how it works. Debian doesn't review software. We package what upstream provides and patch what we distribute. You should be upstream to Debian and Fedora and other distros which are too numerous to mention. Please post your source code online if you want to build an active community around improving it. Software always has room for improvement and never enough people to contribute. I feel it's optimal to bring more people onboard earlier - use the 80:20 rule to decide when to stop going it alone. By asking for review now, you appear to have reached that point, so the time has come to open-source your code for public comment and improvement, and yes, incorporation into mobile phone software without your further permission. It depends on what license you use, but be assured it won't find a place in the Debian distro without open-source licensing. Refer to this older Debian blog post about Debian code reviews and what they are meant to accomplish. https://bits.debian.org/2025/02/bits-from-the-dpl-february.html **expose the architecture** If you are ready, book a physical person into a speaker slot at the 2027 Debconf to show us a demo. We like to get together annually to meet with the other volunteers and see interesting stuff. If you land a spot on the main event stage you could have up to 597 people eagerly watching in-person, and more through the live video feed. https://wiki.debian.org/DebConf/27/Venue_Room?action=AttachFile&do=view&target=debconf27_room-assignment_i18n-EN_v2.pdf Without a demo, I can only guess what your software actually does. My suggestion is, resource management architecture could be most impactful in a large application like LibreOffice. Some people use LO-Calc to build enormous spreadsheets containing hundreds of columns and millions of rows (which properly belong in a database, you know). But they manage, somehow. https://trisquel.info/en/forum/how-handle-enormous-files-libreoffice-without-running-out-memory You are more likely to find acceptance for your LLM framework as an extension in an end-user application than making it core to the Linux kernel. If your software replaces a GPU driver, it will become evident through usage whether the best place for it is a kernel module. Robert
