ggml : do not use _GNU_SOURCE gratuitously

[ggerganov]

See ggml-org/whisper.cpp#1027

[ggerganov]

@przemoc

I'm trying to apply the _GNU_SOURCE patch from whisper.cpp but hitting the following error on MacOS:

In file included from llama.cpp:9:
./llama-util.h:189:63: error: use of undeclared identifier 'MADV_WILLNEED'
            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
                                                              ^
./llama-util.h:316:38: error: use of undeclared identifier 'RLIMIT_MEMLOCK'
            if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
                                     ^
2 errors generated.
make: *** [llama.o] Error 1
Error: Process completed with exit code 2.

Any ideas how to resolve?

[przemoc]

Sorry, I will be momentarily afk for rest of the day, so cannot help much.

I would need to look into llama.cpp, but madvise() typically can be replaced with more portable posix_madvise() and MADV_WILLNEED with analogous POSIX_MADV_WILLNEED (unless you sometimes use in the code also advice that is not defined by POSIX.1).

getrlimit() is slightly different case, because you use resource RLIMIT_MEMLOCK that is not defined by POSIX.1 (even if you would look at the latest POSIX.1-2017), it comes from BSD. You may check if having _BSD_SOURCE additionally defined (-D_BSD_SOURCE after already defined -D_XOPEN_SOURCE=600) helps for macOS.

[przemoc]

@ggerganov, see my attempt for llama.cpp in #2035. Please run CI checks there, because for macOS I'm kind of working in the dark. ;)

[przemoc]

As we're dealing with this topic in my PR #2035 for llama.cpp, I think you may close your own PR #1992 (that was based on my initial attempt at whisper.cpp) to declutter PR space. :)