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
|
{
lib,
stdenv,
symlinkJoin,
buildPythonPackage,
fetchFromGitHub,
gitUpdater,
cmake,
# build-system
scikit-build-core,
setuptools,
# dependencies
torch,
scipy,
trove-classifiers,
cudaSupport ? torch.cudaSupport,
cudaPackages ? torch.cudaPackages,
rocmSupport ? torch.rocmSupport,
rocmPackages ? torch.rocmPackages,
rocmGpuTargets ? rocmPackages.clr.localGpuTargets or rocmPackages.clr.gpuTargets,
}:
let
pname = "bitsandbytes";
version = "0.49.1";
brokenConditions = lib.attrsets.filterAttrs (_: cond: cond) {
"CUDA and ROCm are mutually exclusive" = cudaSupport && rocmSupport;
"CUDA is not targeting Linux" = cudaSupport && !stdenv.hostPlatform.isLinux;
};
inherit (cudaPackages) cudaMajorMinorVersion;
rocmMajorMinorVersion = lib.versions.majorMinor rocmPackages.rocm-core.version;
cudaMajorMinorVersionString = lib.replaceStrings [ "." ] [ "" ] cudaMajorMinorVersion;
rocmMajorMinorVersionString = lib.replaceStrings [ "." ] [ "" ] rocmMajorMinorVersion;
# NOTE: torchvision doesn't use cudnn; torch does!
# For this reason it is not included.
cuda-common-redist = with cudaPackages; [
(lib.getDev cuda_cccl) # <thrust/*>
(lib.getDev libcublas) # cublas_v2.h
(lib.getLib libcublas)
(lib.getInclude libcublas) # cublasLt.h
libcurand
libcusolver # cusolverDn.h
(lib.getDev libcusparse) # cusparse.h
(lib.getLib libcusparse) # cusparse.h
(lib.getInclude libcusparse) # cusparse.h
(lib.getDev cuda_cudart) # cuda_runtime.h cuda_runtime_api.h
];
cuda-native-redist = symlinkJoin {
name = "cuda-native-redist-${cudaMajorMinorVersion}";
paths =
with cudaPackages;
[
(lib.getDev cuda_cudart) # cuda_runtime.h cuda_runtime_api.h
(lib.getLib cuda_cudart)
(lib.getStatic cuda_cudart)
cuda_nvcc
]
++ cuda-common-redist;
};
cuda-redist = symlinkJoin {
name = "cuda-redist-${cudaMajorMinorVersion}";
paths = cuda-common-redist;
};
in
buildPythonPackage {
inherit pname version;
pyproject = true;
src = fetchFromGitHub {
owner = "bitsandbytes-foundation";
repo = "bitsandbytes";
tag = version;
hash = "sha256-nNhxDJITXNIZMXuZdzpF5dl1K1kFEVQ0gbTqZnOf/sI=";
};
patches = [
./find-rocm-deps-with-cmake.patch
];
# By default, which library is loaded depends on the result of `torch.cuda.is_available()`.
# When `cudaSupport` is enabled, bypass this check and load the cuda library unconditionally.
# Indeed, in this case, only `libbitsandbytes_cuda124.so` is built. `libbitsandbytes_cpu.so` is not.
# Also, hardcode the path to the previously built library instead of relying on
# `get_cuda_bnb_library_path(cuda_specs)` which relies on `torch.cuda` too.
#
# WARNING: The cuda library is currently named `libbitsandbytes_cudaxxy` for CUDA version `xx.y`
# and `libbitsandbytes_rocmxxy` for ROCm version `xx.y`
# This upstream convention could change at some point and thus break the following patch.
postPatch = (
let
prefix = if cudaSupport then "cuda" else "rocm";
majorMinorVersionString =
if cudaSupport then cudaMajorMinorVersionString else rocmMajorMinorVersionString;
in
lib.optionalString (cudaSupport || rocmSupport) ''
substituteInPlace bitsandbytes/cextension.py \
--replace-fail "if cuda_specs:" "if True:" \
--replace-fail \
"cuda_binary_path = get_cuda_bnb_library_path(cuda_specs)" \
"cuda_binary_path = PACKAGE_DIR / 'libbitsandbytes_${prefix}${majorMinorVersionString}.so'"
''
);
nativeBuildInputs = [
cmake
]
++ lib.optionals cudaSupport [
cudaPackages.cuda_nvcc
]
++ lib.optionals rocmSupport [
rocmPackages.clr
];
build-system = [
scikit-build-core
setuptools
];
buildInputs =
lib.optional cudaSupport cuda-redist
++ lib.optionals rocmSupport (
with rocmPackages;
[
rocm-device-libs
hipblas
rocm-comgr
rocm-runtime
hiprand
rocrand
hipsparse
hipblaslt
rocblas
hipcub
rocprim
]
);
cmakeFlags = [
(lib.cmakeFeature "COMPUTE_BACKEND" (
if cudaSupport then
"cuda"
else if rocmSupport then
"hip"
else
"cpu"
))
]
++ lib.optionals rocmSupport [
# ends up using g++ to build some files it shouldn't
(lib.cmakeFeature "CMAKE_C_COMPILER" "amdclang")
(lib.cmakeFeature "CMAKE_CXX_COMPILER" "amdclang++")
(lib.cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmGpuTargets))
];
CUDA_HOME = lib.optionalString cudaSupport "${cuda-native-redist}";
NVCC_PREPEND_FLAGS = lib.optionals cudaSupport [
"-I${cuda-native-redist}/include"
"-L${cuda-native-redist}/lib"
];
preBuild = ''
make -j $NIX_BUILD_CORES
cd .. # leave /build/source/build
'';
dependencies = [
scipy
torch
trove-classifiers
];
doCheck = false; # tests require CUDA and also GPU access
pythonImportsCheck = [ "bitsandbytes" ];
passthru = {
inherit
cudaSupport
cudaPackages
rocmSupport
rocmPackages
brokenConditions # To help debug when a package is broken due to CUDA support
;
updateScript = gitUpdater {
ignoredVersions = "continuous-release.*";
};
};
meta = {
description = "8-bit CUDA functions for PyTorch";
homepage = "https://github.com/bitsandbytes-foundation/bitsandbytes";
changelog = "https://github.com/bitsandbytes-foundation/bitsandbytes/releases/tag/${version}";
license = lib.licenses.mit;
maintainers = with lib.maintainers; [
bcdarwin
jk
];
};
}
|
