pyrox.dev
Merge pull request #289513 from happysalada/update_llama_cpp
llama-cpp: pull upstream flake changes
+94
-85
pkgs/by-name/ll/llama-cpp/package.nix
+94
-85
pkgs/by-name/ll/llama-cpp/package.nix
···
15
15
, openclSupport ? false
16
16
, clblast
17
17
18
-
, blasSupport ? !rocmSupport && !cudaSupport
19
-
, openblas
18
+
, blasSupport ? builtins.all (x: !x) [ cudaSupport metalSupport openclSupport rocmSupport vulkanSupport ]
20
19
, pkg-config
21
20
, metalSupport ? stdenv.isDarwin && stdenv.isAarch64 && !openclSupport
22
-
, patchelf
23
-
, static ? true # if false will build the shared objects as well
21
+
, vulkanSupport ? false
22
+
, mpiSupport ? false # Increases the runtime closure by ~700M
23
+
, vulkan-headers
24
+
, vulkan-loader
25
+
, ninja
26
+
, git
27
+
, mpi
24
28
}:
25
29
26
30
let
···
28
32
# otherwise we get libstdc++ errors downstream.
29
33
# cuda imposes an upper bound on the gcc version, e.g. the latest gcc compatible with cudaPackages_11 is gcc11
30
34
effectiveStdenv = if cudaSupport then cudaPackages.backendStdenv else stdenv;
35
+
inherit (lib) cmakeBool cmakeFeature optionals;
36
+
37
+
darwinBuildInputs =
38
+
with darwin.apple_sdk.frameworks;
39
+
[
40
+
Accelerate
41
+
CoreVideo
42
+
CoreGraphics
43
+
]
44
+
++ optionals metalSupport [ MetalKit ];
45
+
46
+
cudaBuildInputs = with cudaPackages; [
47
+
cuda_cccl.dev # <nv/target>
48
+
49
+
# A temporary hack for reducing the closure size, remove once cudaPackages
50
+
# have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
51
+
cuda_cudart.dev
52
+
cuda_cudart.lib
53
+
cuda_cudart.static
54
+
libcublas.dev
55
+
libcublas.lib
56
+
libcublas.static
57
+
];
58
+
59
+
rocmBuildInputs = with rocmPackages; [
60
+
clr
61
+
hipblas
62
+
rocblas
63
+
];
64
+
65
+
vulkanBuildInputs = [
66
+
vulkan-headers
67
+
vulkan-loader
68
+
];
31
69
in
32
70
effectiveStdenv.mkDerivation (finalAttrs: {
33
71
pname = "llama-cpp";
34
-
version = "2249";
72
+
version = "2294";
35
73
36
74
src = fetchFromGitHub {
37
75
owner = "ggerganov";
38
76
repo = "llama.cpp";
39
77
rev = "refs/tags/b${finalAttrs.version}";
40
-
hash = "sha256-ikJUToUbA60u/8azR6dPmPyodq/nQe5L2aotlYBclaE=";
78
+
hash = "sha256-uZi4Bj03PgfFV+jS5M+A1sMCWC/GMY5IyyrlR1b4Sh4=";
41
79
};
42
80
43
81
postPatch = ''
···
45
83
--replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
46
84
'';
47
85
48
-
nativeBuildInputs = [ cmake ] ++ lib.optionals blasSupport [ pkg-config ] ++ lib.optionals cudaSupport [
86
+
nativeBuildInputs = [ cmake ninja pkg-config git ]
87
+
++ optionals cudaSupport [
49
88
cudaPackages.cuda_nvcc
50
89
51
90
# TODO: Replace with autoAddDriverRunpath
···
53
92
cudaPackages.autoAddOpenGLRunpathHook
54
93
];
55
94
56
-
buildInputs = lib.optionals effectiveStdenv.isDarwin
57
-
(with darwin.apple_sdk.frameworks; [
58
-
Accelerate
59
-
CoreGraphics
60
-
CoreVideo
61
-
Foundation
62
-
])
63
-
++ lib.optionals metalSupport (with darwin.apple_sdk.frameworks; [
64
-
MetalKit
65
-
])
66
-
++ lib.optionals cudaSupport (with cudaPackages; [
67
-
cuda_cccl.dev # <nv/target>
68
-
69
-
# A temporary hack for reducing the closure size, remove once cudaPackages
70
-
# have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
71
-
cuda_cudart.dev
72
-
cuda_cudart.lib
73
-
cuda_cudart.static
74
-
libcublas.dev
75
-
libcublas.lib
76
-
libcublas.static
77
-
]) ++ lib.optionals rocmSupport [
78
-
rocmPackages.clr
79
-
rocmPackages.hipblas
80
-
rocmPackages.rocblas
81
-
] ++ lib.optionals openclSupport [
82
-
clblast
83
-
] ++ lib.optionals blasSupport [
84
-
openblas
85
-
];
95
+
buildInputs = optionals effectiveStdenv.isDarwin darwinBuildInputs
96
+
++ optionals cudaSupport cudaBuildInputs
97
+
++ optionals mpiSupport mpi
98
+
++ optionals openclSupport [ clblast ]
99
+
++ optionals rocmSupport rocmBuildInputs
100
+
++ optionals vulkanSupport vulkanBuildInputs;
86
101
87
102
cmakeFlags = [
88
-
"-DLLAMA_NATIVE=OFF"
89
-
"-DLLAMA_BUILD_SERVER=ON"
103
+
# -march=native is non-deterministic; override with platform-specific flags if needed
104
+
(cmakeBool "LLAMA_NATIVE" false)
105
+
(cmakeBool "BUILD_SHARED_SERVER" true)
106
+
(cmakeBool "BUILD_SHARED_LIBS" true)
107
+
(cmakeBool "BUILD_SHARED_LIBS" true)
108
+
(cmakeBool "LLAMA_BLAS" blasSupport)
109
+
(cmakeBool "LLAMA_CLBLAST" openclSupport)
110
+
(cmakeBool "LLAMA_CUBLAS" cudaSupport)
111
+
(cmakeBool "LLAMA_HIPBLAS" rocmSupport)
112
+
(cmakeBool "LLAMA_METAL" metalSupport)
113
+
(cmakeBool "LLAMA_MPI" mpiSupport)
114
+
(cmakeBool "LLAMA_VULKAN" vulkanSupport)
90
115
]
91
-
++ lib.optionals metalSupport [
92
-
"-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
93
-
"-DLLAMA_METAL=ON"
94
-
]
95
-
++ lib.optionals cudaSupport [
96
-
"-DLLAMA_CUBLAS=ON"
97
-
]
98
-
++ lib.optionals rocmSupport [
99
-
"-DLLAMA_HIPBLAS=1"
100
-
"-DCMAKE_C_COMPILER=hipcc"
101
-
"-DCMAKE_CXX_COMPILER=hipcc"
102
-
"-DCMAKE_POSITION_INDEPENDENT_CODE=ON"
103
-
]
104
-
++ lib.optionals openclSupport [
105
-
"-DLLAMA_CLBLAST=ON"
106
-
]
107
-
++ lib.optionals blasSupport [
108
-
"-DLLAMA_BLAS=ON"
109
-
"-DLLAMA_BLAS_VENDOR=OpenBLAS"
110
-
]
111
-
++ lib.optionals (!static) [
112
-
(lib.cmakeBool "BUILD_SHARED_LIBS" true)
113
-
];
116
+
++ optionals cudaSupport [
117
+
(
118
+
with cudaPackages.flags;
119
+
cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
120
+
builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
121
+
)
122
+
)
123
+
]
124
+
++ optionals rocmSupport [
125
+
(cmakeFeature "CMAKE_C_COMPILER" "hipcc")
126
+
(cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
114
127
115
-
installPhase = ''
116
-
runHook preInstall
117
-
118
-
mkdir -p $out/bin
119
-
${lib.optionalString (!static) ''
120
-
mkdir $out/lib
121
-
cp libggml_shared.so $out/lib
122
-
cp libllama.so $out/lib
123
-
''}
124
-
125
-
for f in bin/*; do
126
-
test -x "$f" || continue
127
-
${lib.optionalString (!static) ''
128
-
${patchelf}/bin/patchelf "$f" --set-rpath "$out/lib"
129
-
''}
130
-
cp "$f" $out/bin/llama-cpp-"$(basename "$f")"
131
-
done
132
-
133
-
${lib.optionalString metalSupport "cp ./bin/ggml-metal.metal $out/bin/ggml-metal.metal"}
128
+
# Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
129
+
# in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
130
+
# and select the line that matches the current nixpkgs version of rocBLAS.
131
+
# Should likely use `rocmPackages.clr.gpuTargets`.
132
+
"-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
133
+
]
134
+
++ optionals metalSupport [ (cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
135
+
++ optionals blasSupport [ (cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
134
136
135
-
runHook postInstall
137
+
# upstream plans on adding targets at the cmakelevel, remove those
138
+
# additional steps after that
139
+
postInstall = ''
140
+
mv $out/bin/main $out/bin/llama
141
+
mv $out/bin/server $out/bin/llama-server
142
+
mkdir -p $out/include
143
+
cp $src/llama.h $out/include/
136
144
'';
137
145
138
146
passthru.updateScript = nix-update-script {
···
144
152
description = "Port of Facebook's LLaMA model in C/C++";
145
153
homepage = "https://github.com/ggerganov/llama.cpp/";
146
154
license = licenses.mit;
147
-
mainProgram = "llama-cpp-main";
148
-
maintainers = with maintainers; [ dit7ya elohmeier ];
149
-
broken = (effectiveStdenv.isDarwin && effectiveStdenv.isx86_64) || lib.count lib.id [openclSupport blasSupport rocmSupport cudaSupport] == 0;
155
+
mainProgram = "llama";
156
+
maintainers = with maintainers; [ dit7ya elohmeier philiptaron ];
150
157
platforms = platforms.unix;
158
+
badPlatforms = optionals (cudaSupport || openclSupport) lib.platforms.darwin;
159
+
broken = (metalSupport && !effectiveStdenv.isDarwin);
151
160
};
152
161
})