gpo.zugaina.org

Search Portage & Overlays:

app-local-ai/ik-llama-cpp

LocalAI text-generation backend (ik_llama.cpp gRPC server)

Screenshots

  • ik-llama-cpp-4.10.0
    ~amd64
    +amdgpu_targets_gfx908 +amdgpu_targets_gfx90a +amdgpu_targets_gfx942 +amdgpu_targets_gfx950 +amdgpu_targets_gfx1030 +amdgpu_targets_gfx1100 +amdgpu_targets_gfx1101 +amdgpu_targets_gfx1200 +amdgpu_targets_gfx1201 amdgpu_targets_gfx803 amdgpu_targets_gfx900 amdgpu_targets_gfx906 amdgpu_targets_gfx940 amdgpu_targets_gfx941 amdgpu_targets_gfx1010 amdgpu_targets_gfx1011 amdgpu_targets_gfx1012 amdgpu_targets_gfx1031 amdgpu_targets_gfx1102 amdgpu_targets_gfx1103 amdgpu_targets_gfx1150 amdgpu_targets_gfx1151 cuda native openblas rocm vulkan video_cards_amdgpu

    View      Download      Browse     License: MIT   
    Overlay: local-ai

ChangeLog

commit 81aa9b3707336ed9376007ee763571065779dbb0
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Sat Sep 26 13:55:50 2026 +0000

ik-llama-cpp: create the glue dir prepare.sh expects

The fork's legacy prepare.sh, unlike mainline llama-cpp's, copies into
examples/grpc-server without creating it - upstream's Makefile mkdirs
it before calling the script. Never released, fixed in place.

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit da5a0c0a30c18181d2f9546e10281ec7420ccf84
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Wed Sep 23 09:33:44 2026 +0000

ik-llama-cpp: new backend, ik_llama.cpp gRPC server at the 4.10.0 pin

New app-local-ai/ik-llama-cpp, LocalAI's ik-llama-cpp backend
(ikawrakow's llama.cpp fork: IQ_K quants, faster CPU/hybrid inference)
built like llama-cpp: upstream prepare.sh assembles the glue into the
engine tree, only the grpc-server target is built.

Deltas from the llama-cpp ebuild:
- LOCAL_AI_HIP_CMAKE_VARS=GGML_HIPBLAS: the fork predates the GGML_HIP
rename; the eclass default would silently drop ROCm
- glue lives in examples/ (legacy layout), has no test suite
- no curl: upstream builds with LLAMA_CURL=OFF
- upstream ships this backend CPU-only; the rocm/vulkan USE flags
exceed upstream and are untested there

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit 5ff1d1ea91153421a40a69ae49fcea534cd4baf5
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Fri Aug 28 10:03:52 2026 +0000

drop ik-llama-cpp: HIP port broken at the pin and no model depends on it

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit 1216f2a574c014b304615ea58b4cff433f3277cb
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Fri Aug 28 08:05:23 2026 +0000

define the CUDA-branch tunables for ik-llama-cpp ROCm compiles

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit 0a845e59da65a0d556b18d55b2cebc113c990ff7
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Fri Aug 28 07:49:20 2026 +0000

use the fork's GGML_HIPBLAS toggle for ik-llama-cpp ROCm

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit 274d31c1fb243b6de7c5ff62442cdd2f700efdcf
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Fri Aug 28 07:44:59 2026 +0000

create grpc-server glue dir before ik-llama-cpp prepare.sh

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>

commit 4a4bba2bf70ac2b5330e1523a147e7f3bf9b5c06
Author: Plamen K. Kosseff <plamen@ipnmod.org>
Date: Fri Aug 28 07:21:02 2026 +0000

add ik-llama-cpp and backends-meta, drop 4.8.2, split ggml eclass hierarchy, maintainer metadata

Assisted-by: Claude:claude-fable-5
Signed-off-by: Plamen K. Kosseff <plamen@ipnmod.org>