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
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
|
{ lib
, buildGo122Module
, fetchFromGitHub
, fetchpatch
, buildEnv
, linkFarm
, overrideCC
, makeWrapper
, stdenv
, cmake
, gcc12
, clblast
, libdrm
, rocmPackages
, cudaPackages
, linuxPackages
, darwin
, nixosTests
, testers
, ollama
, ollama-rocm
, ollama-cuda
, config
# one of `[ null false "rocm" "cuda" ]`
, acceleration ? null
}:
let
pname = "ollama";
# don't forget to invalidate all hashes each update
version = "0.1.42";
src = fetchFromGitHub {
owner = "ollama";
repo = "ollama";
rev = "v${version}";
hash = "sha256-lmnfFJBPgjaCdxkMALNQigrtD/V2T3Vs1GEKvRCgWaM=";
fetchSubmodules = true;
};
vendorHash = "sha256-LNH3mpxIrPMe5emfum1W10jvXIjKC6GkGcjq1HhpJQo=";
# ollama's patches of llama.cpp's example server
# `ollama/llm/generate/gen_common.sh` -> "apply temporary patches until fix is upstream"
# each update, these patches should be synchronized with the contents of `ollama/llm/patches/`
llamacppPatches = [
(preparePatch "01-load-progress.diff" "sha256-3QxyKX1n5NeMLU8d7wI/96wCM1Cvb5X5sQL5CFhMFo4=")
(preparePatch "02-clip-log.diff" "sha256-rMWbl3QgrPlhisTeHwD7EnGRJyOhLB4UeS7rqa0tdXM=")
(preparePatch "03-load_exception.diff" "sha256-0XfMtMyg17oihqSFDBakBtAF0JwhsR188D+cOodgvDk=")
(preparePatch "04-metal.diff" "sha256-Ne8J9R8NndUosSK0qoMvFfKNwqV5xhhce1nSoYrZo7Y=")
(preparePatch "05-default-pretokenizer.diff" "sha256-NrQ0Fv5DAZYtRM0NBEeM2JLVTLFmb4Fs9RhwXhdMCC4=")
(preparePatch "06-qwen2.diff" "sha256-nMtoAQUsjYuJv45uTlz8r/K1oF5NUsc75SnhgfSkE30=")
];
preparePatch = patch: hash: fetchpatch {
url = "file://${src}/llm/patches/${patch}";
inherit hash;
stripLen = 1;
extraPrefix = "llm/llama.cpp/";
};
accelIsValid = builtins.elem acceleration [ null false "rocm" "cuda" ];
validateFallback = lib.warnIf (config.rocmSupport && config.cudaSupport)
(lib.concatStrings [
"both `nixpkgs.config.rocmSupport` and `nixpkgs.config.cudaSupport` are enabled, "
"but they are mutually exclusive; falling back to cpu"
])
(!(config.rocmSupport && config.cudaSupport));
shouldEnable = assert accelIsValid;
mode: fallback:
(acceleration == mode)
|| (fallback && acceleration == null && validateFallback);
rocmRequested = shouldEnable "rocm" config.rocmSupport;
cudaRequested = shouldEnable "cuda" config.cudaSupport;
enableRocm = rocmRequested && stdenv.isLinux;
enableCuda = cudaRequested && stdenv.isLinux;
rocmLibs = [
rocmPackages.clr
rocmPackages.hipblas
(rocmPackages.rocblas.override { tensileSepArch = true; tensileLazyLib = true; })
rocmPackages.rocsolver
rocmPackages.rocsparse
rocmPackages.rocm-device-libs
rocmPackages.rocm-smi
];
rocmClang = linkFarm "rocm-clang" {
llvm = rocmPackages.llvm.clang;
};
rocmPath = buildEnv {
name = "rocm-path";
paths = rocmLibs ++ [ rocmClang ];
};
cudaToolkit = buildEnv {
name = "cuda-toolkit";
ignoreCollisions = true; # FIXME: find a cleaner way to do this without ignoring collisions
paths = [
cudaPackages.cudatoolkit
cudaPackages.cuda_cudart
cudaPackages.cuda_cudart.static
];
};
runtimeLibs = lib.optionals enableRocm [
rocmPath
] ++ lib.optionals enableCuda [
linuxPackages.nvidia_x11
];
appleFrameworks = darwin.apple_sdk_11_0.frameworks;
metalFrameworks = [
appleFrameworks.Accelerate
appleFrameworks.Metal
appleFrameworks.MetalKit
appleFrameworks.MetalPerformanceShaders
];
goBuild =
if enableCuda then
buildGo122Module.override { stdenv = overrideCC stdenv gcc12; }
else
buildGo122Module;
inherit (lib) licenses platforms maintainers;
in
goBuild ((lib.optionalAttrs enableRocm {
ROCM_PATH = rocmPath;
CLBlast_DIR = "${clblast}/lib/cmake/CLBlast";
}) // (lib.optionalAttrs enableCuda {
CUDA_LIB_DIR = "${cudaToolkit}/lib";
CUDACXX = "${cudaToolkit}/bin/nvcc";
CUDAToolkit_ROOT = cudaToolkit;
}) // {
inherit pname version src vendorHash;
nativeBuildInputs = [
cmake
] ++ lib.optionals enableRocm [
rocmPackages.llvm.bintools
] ++ lib.optionals (enableRocm || enableCuda) [
makeWrapper
] ++ lib.optionals stdenv.isDarwin
metalFrameworks;
buildInputs = lib.optionals enableRocm
(rocmLibs ++ [ libdrm ])
++ lib.optionals enableCuda [
cudaPackages.cuda_cudart
] ++ lib.optionals stdenv.isDarwin
metalFrameworks;
patches = [
# disable uses of `git` in the `go generate` script
# ollama's build script assumes the source is a git repo, but nix removes the git directory
# this also disables necessary patches contained in `ollama/llm/patches/`
# those patches are added to `llamacppPatches`, and reapplied here in the patch phase
./disable-git.patch
# disable a check that unnecessarily exits compilation during rocm builds
# since `rocmPath` is in `LD_LIBRARY_PATH`, ollama uses rocm correctly
./disable-lib-check.patch
] ++ llamacppPatches;
postPatch = ''
# replace inaccurate version number with actual release version
substituteInPlace version/version.go --replace-fail 0.0.0 '${version}'
'';
preBuild = ''
# disable uses of `git`, since nix removes the git directory
export OLLAMA_SKIP_PATCHING=true
# build llama.cpp libraries for ollama
go generate ./...
'';
postFixup = ''
# the app doesn't appear functional at the moment, so hide it
mv "$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" ${
lib.optionalString enableRocm
''--set-default HIP_PATH '${rocmPath}' ''} \
--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;
version = testers.testVersion {
inherit version;
package = ollama;
};
} // stdenv.isLinux {
inherit ollama-rocm ollama-cuda;
};
meta = {
description = "Get up and running with large language models locally"
+ lib.optionalString rocmRequested ", using ROCm for AMD GPU acceleration"
+ lib.optionalString cudaRequested ", using CUDA for NVIDIA GPU acceleration";
homepage = "https://github.com/ollama/ollama";
changelog = "https://github.com/ollama/ollama/releases/tag/v${version}";
license = licenses.mit;
platforms =
if (rocmRequested || cudaRequested) then platforms.linux
else platforms.unix;
mainProgram = "ollama";
maintainers = with maintainers; [ abysssol dit7ya elohmeier roydubnium ];
};
})
|
