kvmem-cuda-git
LOW
maintainer chrishamm
0 votes
scanned 2026-11-01 16:01:05.312492
Why flagged
Uploaded within the last 14 days with 2 or fewer community votes — little peer review so far.
Triggered rules
Low
Few votes, recently uploaded
zero_votes_recent
Uploaded within the last 14 days with 2 or fewer community votes — little peer review so far.
PKGBUILD
1
# Maintainer: Christian Hammacher <chr.hammacher@gmail.com>
2
# Based on the extra/llama-cpp PKGBUILD by Sven-Hendrik Haase and Maxime Gauduin
3
4
pkgname=kvmem-cuda-git
5
pkgver=0.19.1.r20.g4b755f2
6
pkgrel=1
7
pkgdesc='llama.cpp server with KVMem long-context KV memory management (CUDA)'
8
arch=(x86_64)
9
url=https://github.com/kvmem/kvmem
10
license=(
11
MIT
12
LicenseRef-unknown
13
)
14
depends=(
15
cuda
16
glibc
17
libgcc
18
libstdc++
19
nvidia-utils
20
openssl
21
)
22
makedepends=(
23
chrpath
24
cmake
25
git
26
ninja
27
python
28
)
29
provides=(kvmem)
30
conflicts=(kvmem)
31
options=(
32
!buildflags
33
!debug
34
)
35
source=(git+https://github.com/kvmem/kvmem.git
36
git+https://github.com/ggml-org/llama.cpp.git
37
0001-Keep-nvml.h-from-clobbering-llama.h-DEPRECATED.patch)
38
b2sums=('SKIP'
39
'SKIP'
40
'c1adf0f5dd86209807ac7f2fb45045dc911bc98e6f9cdeeb113ed7724eed857ad6bc30bdca3c4ee69c752affdd0079af5dc1d4816fb1bcc74a52ad22c1677553')
41
42
# The adapter patches llama.cpp's bundled ggml, so neither system ggml nor
43
# llama-cpp can be used; the private libraries live in /usr/lib/kvmem instead
44
45
# RTX 20/30/40/50, the cards KVMem qualifies; set CUDA_ARCHITECTURES to build for one card only
46
_cuda_arch=${CUDA_ARCHITECTURES:-75-real;86-real;89-real;120a-real}
47
48
pkgver() {
49
cd kvmem
50
git describe --long --tags --abbrev=7 | sed 's/^v//;s/\([^-]*-g\)/r\1/;s/-/./g'
51
}
52
53
prepare() {
54
cd kvmem
55
git submodule init backends/llamacpp
56
git config submodule.llamacpp.url "${srcdir}"/llama.cpp
57
git -c protocol.file.allow=always submodule update --no-recommend-shallow backends/llamacpp
58
python scripts/prepare-backends.py --backend llamacpp
59
patch -Np1 -i ../0001-Keep-nvml.h-from-clobbering-llama.h-DEPRECATED.patch
60
}
61
62
build() {
63
cmake -S kvmem -B build -G Ninja \
64
-DCMAKE_BUILD_TYPE=Release \
65
-DCMAKE_CUDA_COMPILER=/opt/cuda/bin/nvcc \
66
-DCMAKE_CUDA_ARCHITECTURES="${_cuda_arch}" \
67
-DGGML_NATIVE=OFF \
68
-DGGML_CUDA=ON \
69
-DGGML_CUDA_FA_QUANTS=all \
70
-DKVMEM_BUILD_LLAMA=ON \
71
-DKVMEM_BUILD_SERVER_TESTS=OFF \
72
-DLLAMA_KVMEM=ON \
73
-DLLAMA_KVMEM_ROOT="${srcdir}"/kvmem
74
cmake --build build --target llama-kvmem-server
75
}
76
77
package() {
78
install -Dm 755 build/bin/llama-kvmem-server -t "${pkgdir}"/usr/bin/
79
install -dm 755 "${pkgdir}"/usr/lib/kvmem
80
cp -a build/bin/lib{ggml,ggml-base,ggml-cpu,ggml-cuda,llama,llama-common,mtmd}.so* "${pkgdir}"/usr/lib/kvmem/
81
local _f
82
for _f in "${pkgdir}"/usr/bin/llama-kvmem-server $(find "${pkgdir}"/usr/lib/kvmem -type f); do
83
if chrpath -l "${_f}" >/dev/null; then
84
chrpath -r /usr/lib/kvmem "${_f}"
85
fi
86
done
87
install -Dm 644 kvmem/backends/llamacpp/LICENSE -t "${pkgdir}"/usr/share/licenses/${pkgname}/
88
}
89
Scan history
| Scanned at (UTC) | Severity | Rules |
|---|---|---|
| 2026-11-01 16:01:05 | Low | 1 |