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
223
224
225
226
227
228
229
230
231
232
233
234
|
{ lib
, buildGoModule
, fetchFromGitHub
, fetchpatch
, buildEnv
, linkFarm
, overrideCC
, makeWrapper
, stdenv
, addDriverRunpath
, cmake
, gcc12
, clblast
, libdrm
, rocmPackages
, cudaPackages
, darwin
, autoAddDriverRunpath
, 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 = "0.2.7";
src = fetchFromGitHub {
owner = "ollama";
repo = "ollama";
rev = "v${version}";
hash = "sha256-YHBGS615dxz6pZR+8awMinp+ZRf3J8Re5BVeDEIyt2E=";
fetchSubmodules = true;
};
vendorHash = "sha256-hSxcREAujhvzHVNwnRTfhi0MKI3s8HNavER2VLz6SYk=";
# ollama's patches of llama.cpp's example server
# `ollama/llm/generate/gen_common.sh` -> "apply temporary patches until fix is upstream"
# each update, these patches should be synchronized with the contents of `ollama/llm/patches/`
llamacppPatches = [
(preparePatch "01-load-progress.diff" "sha256-UTmnBS5hQjIL3eXDZc8RBDNJunLlkqJWH20LpXNiGRQ=")
(preparePatch "02-clip-log.diff" "sha256-rMWbl3QgrPlhisTeHwD7EnGRJyOhLB4UeS7rqa0tdXM=")
(preparePatch "03-load_exception.diff" "sha256-NJkT/k8Mf8HcEMb0XkaLmyUNKV3T+384JRPnmwDI/sk=")
(preparePatch "04-metal.diff" "sha256-bPBCfoT3EjZPjWKfCzh0pnCUbM/fGTj37yOaQr+QxQ4=")
(preparePatch "05-default-pretokenizer.diff" "sha256-50+mzQBQZmYEhYvARHw/dliH0M/gDOYm2uy/yJupDF4=")
(preparePatch "06-qwen2.diff" "sha256-FdDqEIblPy47z3yavKUnaV93Yk+3oboEzj4vHq+R66M=")
(preparePatch "07-embeddings.diff" "sha256-lqg2SI0OapD9LCoAG6MJW6HIHXEmCTv7P75rE9yq/Mo=")
(preparePatch "08-clip-unicode.diff" "sha256-1qMJoXhDewxsqPbmi+/7xILQfGaybZDyXc5eH0winL8=")
(preparePatch "09-pooling.diff" "sha256-7meKWbr06lbVrtxau0AU9BwJ88Z9svwtDXhmHI+hYBk=")
];
preparePatch = patch: hash: fetchpatch {
url = "file://${src}/llm/patches/${patch}";
inherit hash;
stripLen = 1;
extraPrefix = "llm/llama.cpp/";
};
accelIsValid = builtins.elem acceleration [ null false "rocm" "cuda" ];
validateFallback = lib.warnIf (config.rocmSupport && config.cudaSupport)
(lib.concatStrings [
"both `nixpkgs.config.rocmSupport` and `nixpkgs.config.cudaSupport` are enabled, "
"but they are mutually exclusive; falling back to cpu"
])
(!(config.rocmSupport && config.cudaSupport));
shouldEnable = assert accelIsValid;
mode: fallback:
(acceleration == mode)
|| (fallback && acceleration == null && validateFallback);
rocmRequested = shouldEnable "rocm" config.rocmSupport;
cudaRequested = shouldEnable "cuda" config.cudaSupport;
enableRocm = rocmRequested && stdenv.isLinux;
enableCuda = cudaRequested && stdenv.isLinux;
rocmLibs = [
rocmPackages.clr
rocmPackages.hipblas
rocmPackages.rocblas
rocmPackages.rocsolver
rocmPackages.rocsparse
rocmPackages.rocm-device-libs
rocmPackages.rocm-smi
];
rocmClang = linkFarm "rocm-clang" {
llvm = rocmPackages.llvm.clang;
};
rocmPath = buildEnv {
name = "rocm-path";
paths = rocmLibs ++ [ rocmClang ];
};
cudaToolkit = buildEnv {
name = "cuda-merged";
paths = [
(lib.getBin (cudaPackages.cuda_nvcc.__spliced.buildHost or cudaPackages.cuda_nvcc))
(lib.getLib cudaPackages.cuda_cudart)
(lib.getOutput "static" cudaPackages.cuda_cudart)
(lib.getLib cudaPackages.libcublas)
];
};
appleFrameworks = darwin.apple_sdk_11_0.frameworks;
metalFrameworks = [
appleFrameworks.Accelerate
appleFrameworks.Metal
appleFrameworks.MetalKit
appleFrameworks.MetalPerformanceShaders
];
wrapperOptions = [
# ollama embeds llama-cpp binaries which actually run the ai models
# these llama-cpp binaries are unaffected by the ollama binary's DT_RUNPATH
# LD_LIBRARY_PATH is temporarily required to use the gpu
# until these llama-cpp binaries can have their runpath patched
"--suffix LD_LIBRARY_PATH : '${addDriverRunpath.driverLink}/lib'"
] ++ lib.optionals enableRocm [
"--suffix LD_LIBRARY_PATH : '${rocmPath}/lib'"
"--set-default HIP_PATH '${rocmPath}'"
];
wrapperArgs = builtins.concatStringsSep " " wrapperOptions;
goBuild =
if enableCuda then
buildGoModule.override { stdenv = overrideCC stdenv gcc12; }
else
buildGoModule;
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";
}) // {
inherit pname version src vendorHash;
nativeBuildInputs = [
cmake
] ++ lib.optionals enableRocm [
rocmPackages.llvm.bintools
] ++ lib.optionals enableCuda [
cudaPackages.cuda_nvcc
] ++ lib.optionals (enableRocm || enableCuda) [
makeWrapper
autoAddDriverRunpath
] ++ lib.optionals stdenv.isDarwin
metalFrameworks;
buildInputs = lib.optionals enableRocm
(rocmLibs ++ [ libdrm ])
++ lib.optionals enableCuda [
cudaPackages.cuda_cudart
cudaPackages.cuda_cccl
cudaPackages.libcublas
] ++ 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" "$out/bin/.ollama-app"
'' + lib.optionalString (enableRocm || enableCuda) ''
# expose runtime libraries necessary to use the gpu
wrapProgram "$out/bin/ollama" ${wrapperArgs}
'';
ldflags = [
"-s"
"-w"
"-X=github.com/ollama/ollama/version.Version=${version}"
"-X=github.com/ollama/ollama/server.mode=release"
];
passthru.tests = {
inherit ollama;
service = nixosTests.ollama;
version = testers.testVersion {
inherit version;
package = ollama;
};
} // lib.optionalAttrs stdenv.isLinux {
inherit ollama-rocm ollama-cuda;
service-cuda = nixosTests.ollama-cuda;
service-rocm = nixosTests.ollama-rocm;
};
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 ];
};
})
|
