Ok I think I see a few options here:

1. We fork whisper.cpp as a separate repo, apply the patches there, and then pull it in to this project as part of a setup script. No git submodule, the dir it gets pulled into is git ignored
2. We use the design of this PR where all the patches exist and are applied in the llamafile repo itself.
3. Use a build system like Conan or cmake. I tend towards not wanting this rn since it'll add a little more tooling overhead but idk

Since the patches currently involve pulling in llamafile dependencies to whisper, option 2 seems like the easiest lift that would bring organizational improvements.

After discussion with @aittalam. we think the git submodule approach will give us a good base to split out third party logic from our codebase, so that we can have a good base to make informed upgrade decisions

Thanks @njbrake! I replicated this and verified it regenerates the same code we had before, while making changes more apparent. I think it is gonna be helpful to better understand what changed, how, and how to align the changes with more recent versions of the code.

I left a couple of questions / fixes, plus one more open one here: what is your take on bringing all customised files to the same dir vs keeping their original structure? I do not have strong opinions: I originally had the intuition that subdirs might make it easier to compare with original code, but now that we have patches for that (and an explicit mapping in the apply-patches.sh script) we probably do not need that much anymore.

For the moment, I think I would like to keep their original structure. In future PRs we can continue clean up to revert them to their original structure, but I wanted to keep this PR as simple as possible: to keep the risk of the PR at the lowest possible, I wanted this PR to only switch to using git submodules, and in future PRs we can consider directory restructuring.

For the moment, I think I would like to keep their original structure. In future PRs we can continue clean up to revert them to their original structure, but I wanted to keep this PR as simple as possible: to keep the risk of the PR at the lowest possible, I wanted this PR to only switch to using git submodules, and in future PRs we can consider directory restructuring.

Ah yes, completely agree with it for now! Our target ATM is to replicate exactly what we have, and just get more signal re: the changes