xdelta3 patch files and less frequent builds

[zmn28hgbn59kcmlpio8unfh7fdre523esd28q9a]

hello,

llama.cpp can have multiple builds per day, and Windows builds are the largest; for most people who want to be up-to-date are fine with that, however some ISPs are not that generous with bandwidth with some significantly reduce bandwidth after some threshold or worse they block internet access.

xdelta3 is binary diff tool that create very small patch files, and is available on all major platform. (Windows, FreeBSD, Linux and OS X)

the downside of using patch files:

• if per file patching is used, then it is tedious depending on how the user is applying the patches
• per archive, e.g. llama--bin-win-cuda-12.4-x64.zip, then the original archive must be kept and renamed if necessary to match newest version before applying a new patch (xdelta store file names and checksum in .patch file)

the thing that renders patch files useless is llama.cpp can have more than 5 builds per day, and it is very easy not to keep up with llama.cpp builds per day, and if care is not taken, say for example a user was busy with life for one or two days, then the user will download the latest version without hesitation because nobody wants to download dozens of patch files, which in turn makes the idea of patch files completely waste of time and energy for llama.cpp team.

the good thing about patch file is obviously reduced bandwidth provided that a user have copy of the previous build for example.

DO NOT consider this terrible idea, since it doesnt line up with llama.cpp release schedule, in either cases i just wanted to drop it here because my head is spinning of how many times i see some good additions and fixes in newer llama.cpp builds.

thank you.