Re: llama.cpp, whisper.cpp, ggml: Next steps

[Christian Kastner <ckk@debian.org>]

On 2025-02-27 23:12, Petter Reinholdtsen wrote:
> [Christian Kastner]
> Thanks.  Tried this approach, and get the same result.  But the binaries
> are missing the required rpath/runpath:
>  
>   % ldd debian/whisper.cpp/usr/bin/whisper-stream |grep not 
>           libggml.so => not found
>           libggml-base.so => not found

Hm, that's odd for many reasons. For one, whisper-stream shouldn't need
ggml, ldd is showing you libwhisper1's transitive dependency.

Is libggml installed in the environment you tested this in? If not, that
would explain the "not found".

Just looking at whisper-stream:

 $ objdump -x whisper-stream | grep -E '(NEEDED|RUNPATH)'
  NEEDED               libwhisper.so.1
  NEEDED               libSDL2-2.0.so.0
  NEEDED               libstdc++.so.6
  NEEDED               libm.so.6
  NEEDED               libgcc_s.so.1
  NEEDED               libc.so.6
  RUNPATH              /usr/lib/x86_64-linux-gnu/ggml

So (1) no ggml, but (2) RUNPATH *is* set (which we might actually not
want for the binaries).

I also don't see any direct references to ggml in examples/stream, which
is what I would expect.

>> By the way -- why TARGET instead of HOST? Genuinely curious, I thought
>> its use case was just for toolchains.
> 
> Which TARGET?  Most likely it is because I have no idea how to properly
> adjust cmake and is just cargo-culting until I get something limping
> along. :)

Oops, that was ambiguous. I meant: why DEB_TARGET_MULTIARCH instead of
the more common DEB_HOST_MULTIARCH.

I don't think I've encountered DEB_TARGET_ARCH before, so I may be
overlooking something when using _HOST_.

Best,
Christian