Arch Linux User Repository

Alternate versions

llama.cpp
llama.cpp-vulkan
llama.cpp-sycl-fp16
llama.cpp-sycl-fp32
llama.cpp-cuda
llama.cpp-cuda-f16
llama.cpp-hip

@kelvie

1. It doesn't actually use it, it's just downloaded for no reason (this package is the CPU backend). The nomic-ai fork is what llama.cpp uses for its Kompute backend (GGML_KOMPUTE=ON).
2. CMAKE_BUILD_TYPE=None is the standard for building Arch packages. See the CMake package guidelines page in the wiki: makepkg provides its own release flags.

(llama.cpp assertions via GGML_ASSERT are always enabled in all build types.)

2 questions:

1. Why does this use a strange fork of kompute?
2. Why is CMAKE_BUILD_TYPE set to none rather than Release? I think this leaves asserts enabled.

@cfillion I got rid of OpenBLAS. Please let me know if it's working as expected.

Fair enough. Is OpenBLAS actually faster for most users though? It's unbearably slow here.

export TEXT="Arch Linux defines simplicity as without unnecessary additions or modifications. It ships software as released by the original developers (upstream) with minimal distribution-specific (downstream) changes: patches not accepted by upstream are avoided, and Arch's downstream patches consist almost entirely of backported bug fixes that are obsoleted by the project's next release."

# GGML_BLAS=ON GGML_BLAS_VENDOR=OpenBLAS
$ time llama-cli -m Mistral-Small-Instruct-2409-Q6_K_L.gguf -tb 16 -p "$TEXT" -n 1 -no-cnv
llama_perf_context_print: prompt eval time =   10825.25 ms /    76 tokens (  142.44 ms per token,     7.02 tokens per second)
291.99s user 1.91s system 2544% cpu 11.551 total

# GGML_BLAS=OFF
$ time llama-cli -m Mistral-Small-Instruct-2409-Q6_K_L.gguf -tb 16 -p "$TEXT" -n 1 -no-cnv
llama_perf_context_print: prompt eval time =    2248.24 ms /    76 tokens (   29.58 ms per token,    33.80 tokens per second)
40.49s user 0.54s system 1395% cpu 2.940 total

# GGML_BLAS=OFF (w/ multiprocessing to imitate OpenBLAS's behavior, ruling it out as a single cause)
$ time llama-cli -m Mistral-Small-Instruct-2409-Q6_K_L.gguf -tb 32 -p "$TEXT" -n 1 -no-cnv
llama_perf_context_print: prompt eval time =    2350.10 ms /    76 tokens (   30.92 ms per token,    32.34 tokens per second)
85.28s user 0.78s system 2784% cpu 3.091 total

@cfillion I'm already maintaining several variants of this package. I don't think I'll add another variant unless it's popularly requested.

Thanks for the feedback though!

GGML_BLAS=ON with OpenBLAS is significantly slower at prompt processing on my CPU (9950X w/ n_threads = n_threads_batch = 16) than OFF. Could there be a variant of the package without it?

Also, isn't the kompute submodule unnecessary/unused unless GLML_KOMPUTE=ON is set? (which would be for a separate llama.cpp-kompute package?)

@abitrolly Because of the maintenance overhead. It's easier to main as separate packages.

@txtid why one source package can not produce all these binaries? Why the binaries can not be renamed to avoid conflict?

@abitrolly The split packages all conflict with each other. You can only have one installed at any point, so it doesn't make sense to build all just to get one. Additionally, the -hip and -cuda packages take very long to build.

I don't think anyone wants to install the 20GB+ dependencies and compile all variants just to get one part of the split package.

@txtsd actually it would be nice to get everything what is necessary to get the most optimized build for a system in one go.