1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
|
{ lib
, buildGoModule
, fetchFromGitHub
, fetchpatch
, buildEnv
, linkFarm
, overrideCC
, makeWrapper
, stdenv
, nixosTests
, pkgs
, cmake
, gcc12
, clblast
, libdrm
, rocmPackages
, cudaPackages
, linuxPackages
, darwin
# one of `[ null "rocm" "cuda" ]`
, acceleration ? null
}:
let
pname = "ollama";
version = "0.1.28";
src = fetchFromGitHub {
owner = "jmorganca";
repo = "ollama";
rev = "v${version}";
hash = "sha256-8f7veZitorNiqGBPJuf/Y36TcFK8Q75Vw4w6CeTk8qs=";
fetchSubmodules = true;
};
validAccel = lib.assertOneOf "ollama.acceleration" acceleration [ null "rocm" "cuda" ];
warnIfNotLinux = api: (lib.warnIfNot stdenv.isLinux
"building ollama with `${api}` is only supported on linux; falling back to cpu"
stdenv.isLinux);
enableRocm = validAccel && (acceleration == "rocm") && (warnIfNotLinux "rocm");
enableCuda = validAccel && (acceleration == "cuda") && (warnIfNotLinux "cuda");
rocmClang = linkFarm "rocm-clang" {
llvm = rocmPackages.llvm.clang;
};
rocmPath = buildEnv {
name = "rocm-path";
paths = [
rocmPackages.rocm-device-libs
rocmClang
];
};
cudaToolkit = buildEnv {
name = "cuda-toolkit";
ignoreCollisions = true; # FIXME: find a cleaner way to do this without ignoring collisions
paths = [
cudaPackages.cudatoolkit
cudaPackages.cuda_cudart
];
};
runtimeLibs = lib.optionals enableRocm [
rocmPackages.rocm-smi
] ++ lib.optionals enableCuda [
linuxPackages.nvidia_x11
];
appleFrameworks = darwin.apple_sdk_11_0.frameworks;
metalFrameworks = [
appleFrameworks.Accelerate
appleFrameworks.Metal
appleFrameworks.MetalKit
appleFrameworks.MetalPerformanceShaders
];
goBuild =
if enableCuda then
buildGoModule.override { stdenv = overrideCC stdenv gcc12; }
else
buildGoModule;
preparePatch = patch: hash: fetchpatch {
url = "file://${src}/llm/patches/${patch}";
inherit hash;
stripLen = 1;
extraPrefix = "llm/llama.cpp/";
};
inherit (lib) licenses platforms maintainers;
in
goBuild ((lib.optionalAttrs enableRocm {
ROCM_PATH = rocmPath;
CLBlast_DIR = "${clblast}/lib/cmake/CLBlast";
}) // (lib.optionalAttrs enableCuda {
CUDA_LIB_DIR = "${cudaToolkit}/lib";
CUDACXX = "${cudaToolkit}/bin/nvcc";
CUDAToolkit_ROOT = cudaToolkit;
}) // {
inherit pname version src;
vendorHash = "sha256-DPIhDqE/yXpSQqrx07osMBMafK61yU2dl4cZhxSTvm8=";
nativeBuildInputs = [
cmake
] ++ lib.optionals (enableRocm || enableCuda) [
makeWrapper
] ++ lib.optionals stdenv.isDarwin
metalFrameworks;
buildInputs = lib.optionals enableRocm [
rocmPackages.clr
rocmPackages.hipblas
rocmPackages.rocblas
rocmPackages.rocsolver
rocmPackages.rocsparse
libdrm
] ++ lib.optionals enableCuda [
cudaPackages.cuda_cudart
] ++ lib.optionals stdenv.isDarwin
metalFrameworks;
patches = [
# remove uses of `git` in the `go generate` script
# instead use `patch` where necessary
./remove-git.patch
# replace a hardcoded use of `g++` with `$CXX`
./replace-gcc.patch
# ollama's patches of llama.cpp's example server
# `ollama/llm/generate/gen_common.sh` -> "apply temporary patches until fix is upstream"
(preparePatch "01-cache.diff" "sha256-MTTln2G0G8dntihUzEjPM1ruTsApb4ZToBczJb8EG68=")
(preparePatch "02-cudaleaks.diff" "sha256-Cu7E9iEcvddPL9mPPI5Z96qmwWigi3f0WgSpPRjGc88=")
];
postPatch = ''
# use a patch from the nix store in the `go generate` script
substituteInPlace llm/generate/gen_common.sh \
--subst-var-by cmakeIncludePatch '${./cmake-include.patch}'
# `ollama/llm/generate/gen_common.sh` -> "avoid duplicate main symbols when we link into the cgo binary"
substituteInPlace llm/llama.cpp/examples/server/server.cpp \
--replace-fail 'int main(' 'int __main('
# replace inaccurate version number with actual release version
substituteInPlace version/version.go --replace-fail 0.0.0 '${version}'
'';
preBuild = ''
export OLLAMA_SKIP_PATCHING=true
# build llama.cpp libraries for ollama
go generate ./...
'';
postFixup = ''
# the app doesn't appear functional at the moment, so hide it
mv "$out/bin/app" "$out/bin/.ollama-app"
'' + lib.optionalString (enableRocm || enableCuda) ''
# expose runtime libraries necessary to use the gpu
mv "$out/bin/ollama" "$out/bin/.ollama-unwrapped"
makeWrapper "$out/bin/.ollama-unwrapped" "$out/bin/ollama" \
--suffix LD_LIBRARY_PATH : '/run/opengl-driver/lib:${lib.makeLibraryPath runtimeLibs}'
'';
ldflags = [
"-s"
"-w"
"-X=github.com/jmorganca/ollama/version.Version=${version}"
"-X=github.com/jmorganca/ollama/server.mode=release"
];
passthru.tests = {
service = nixosTests.ollama;
rocm = pkgs.ollama.override { acceleration = "rocm"; };
cuda = pkgs.ollama.override { acceleration = "cuda"; };
};
meta = {
description = "Get up and running with large language models locally";
homepage = "https://github.com/jmorganca/ollama";
license = licenses.mit;
platforms = platforms.unix;
mainProgram = "ollama";
maintainers = with maintainers; [ abysssol dit7ya elohmeier ];
};
})
|