blob: 80dc2b500a0f73001ee7a95ffeac5fecd6264332 (
plain)
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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
|
# Copyright 2022-2024 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2
EAPI=8
PYTHON_COMPAT=( python3_{9..12} )
ROCM_VERSION=5.7
inherit python-single-r1 cmake cuda flag-o-matic prefix rocm
MYPN=pytorch
MYP=${MYPN}-${PV}
DESCRIPTION="A deep learning framework"
HOMEPAGE="https://pytorch.org/"
SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
-> ${MYP}.tar.gz"
LICENSE="BSD"
SLOT="0"
KEYWORDS="~amd64"
IUSE="cuda distributed fbgemm ffmpeg gloo mkl mpi nnpack +numpy onednn openblas opencl opencv openmp qnnpack rocm xnnpack"
RESTRICT="test"
REQUIRED_USE="
${PYTHON_REQUIRED_USE}
ffmpeg? ( opencv )
mpi? ( distributed )
gloo? ( distributed )
?? ( cuda rocm )
rocm? ( || ( ${ROCM_REQUIRED_USE} ) )
"
# CUDA 12 not supported yet: https://github.com/pytorch/pytorch/issues/91122
RDEPEND="
${PYTHON_DEPS}
dev-cpp/gflags:=
>=dev-cpp/glog-0.5.0
dev-libs/cpuinfo
dev-libs/libfmt
dev-libs/protobuf:=
dev-libs/pthreadpool
dev-libs/sleef
virtual/lapack
>=sci-libs/onnx-1.12.0
<sci-libs/onnx-1.15.0
sci-libs/foxi
cuda? (
=dev-libs/cudnn-8*
>=dev-libs/cudnn-frontend-0.9.2:0/8
dev-util/nvidia-cuda-toolkit:=[profiler]
)
fbgemm? ( >=dev-libs/FBGEMM-2023.12.01 )
ffmpeg? ( media-video/ffmpeg:= )
gloo? ( sci-libs/gloo[cuda?] )
mpi? ( virtual/mpi )
nnpack? ( sci-libs/NNPACK )
numpy? ( $(python_gen_cond_dep '
dev-python/numpy[${PYTHON_USEDEP}]
') )
onednn? ( dev-libs/oneDNN )
opencl? ( virtual/opencl )
opencv? ( media-libs/opencv:= )
qnnpack? ( sci-libs/QNNPACK )
rocm? (
>=dev-util/hip-5.7
>=dev-libs/rccl-5.7[${ROCM_USEDEP}]
>=sci-libs/rocThrust-5.7[${ROCM_USEDEP}]
>=sci-libs/rocPRIM-5.7[${ROCM_USEDEP}]
>=sci-libs/hipBLAS-5.7[${ROCM_USEDEP}]
>=sci-libs/hipFFT-5.7[${ROCM_USEDEP}]
>=sci-libs/hipSPARSE-5.7[${ROCM_USEDEP}]
>=sci-libs/hipRAND-5.7[${ROCM_USEDEP}]
>=sci-libs/hipCUB-5.7[${ROCM_USEDEP}]
>=sci-libs/hipSOLVER-5.7[${ROCM_USEDEP}]
>=sci-libs/miopen-5.7[${ROCM_USEDEP}]
>=dev-util/roctracer-5.7[${ROCM_USEDEP}]
)
distributed? ( sci-libs/tensorpipe[cuda?] )
xnnpack? ( >=sci-libs/XNNPACK-2022.12.22 )
mkl? ( sci-libs/mkl )
openblas? ( sci-libs/openblas )
"
DEPEND="
${RDEPEND}
cuda? ( >=dev-libs/cutlass-3.1.0 )
onednn? ( sci-libs/ideep )
dev-libs/psimd
dev-libs/FP16
dev-libs/FXdiv
dev-libs/pocketfft
dev-libs/flatbuffers
>=sci-libs/kineto-0.4.0_p20231031
$(python_gen_cond_dep '
dev-python/pyyaml[${PYTHON_USEDEP}]
dev-python/pybind11[${PYTHON_USEDEP}]
')
"
S="${WORKDIR}"/${MYP}
PATCHES=(
"${FILESDIR}"/${P}-gentoo.patch
"${FILESDIR}"/${PN}-1.13.0-install-dirs.patch
"${FILESDIR}"/${PN}-1.12.0-glog-0.6.0.patch
"${FILESDIR}"/${PN}-1.13.1-tensorpipe.patch
"${FILESDIR}"/${PN}-2.0.0-gcc13.patch
"${FILESDIR}"/${PN}-2.0.0-cudnn_include_fix.patch
"${FILESDIR}"/${PN}-2.1.2-fix-rpath.patch
"${FILESDIR}"/${PN}-2.1.2-fix-openmp-link.patch
"${FILESDIR}"/${PN}-2.1.2-rocm-fix-std-cpp17.patch
)
src_prepare() {
filter-lto #bug 862672
sed -i \
-e "/third_party\/gloo/d" \
cmake/Dependencies.cmake \
|| die
cmake_src_prepare
pushd torch/csrc/jit/serialization || die
flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
popd
# prefixify the hardcoded paths, after all patches are applied
hprefixify \
aten/CMakeLists.txt \
caffe2/CMakeLists.txt \
cmake/Metal.cmake \
cmake/Modules/*.cmake \
cmake/Modules_CUDA_fix/FindCUDNN.cmake \
cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
cmake/public/LoadHIP.cmake \
cmake/public/cuda.cmake \
cmake/Dependencies.cmake \
torch/CMakeLists.txt \
CMakeLists.txt
if use rocm; then
sed -e "s:/opt/rocm:/usr:" \
-e "s:lib/cmake:$(get_libdir)/cmake:g" \
-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
-i cmake/public/LoadHIP.cmake || die
ebegin "HIPifying cuda sources"
${EPYTHON} tools/amd_build/build_amd.py || die
eend $?
fi
}
src_configure() {
if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
ewarn "These may not be optimal for your GPU."
ewarn ""
ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
ewarn ""
ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
fi
local mycmakeargs=(
-DBUILD_CUSTOM_PROTOBUF=OFF
-DBUILD_SHARED_LIBS=ON
-DUSE_CCACHE=OFF
-DUSE_CUDA=$(usex cuda)
-DUSE_DISTRIBUTED=$(usex distributed)
-DUSE_MPI=$(usex mpi)
-DUSE_FAKELOWP=OFF
-DUSE_FBGEMM=$(usex fbgemm)
-DUSE_FFMPEG=$(usex ffmpeg)
-DUSE_GFLAGS=ON
-DUSE_GLOG=ON
-DUSE_GLOO=$(usex gloo)
-DUSE_KINETO=OFF # TODO
-DUSE_LEVELDB=OFF
-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
-DUSE_MKLDNN=$(usex onednn)
-DUSE_NNPACK=$(usex nnpack)
-DUSE_QNNPACK=$(usex qnnpack)
-DUSE_XNNPACK=$(usex xnnpack)
-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
-DUSE_TENSORPIPE=$(usex distributed)
-DUSE_PYTORCH_QNNPACK=OFF
-DUSE_NUMPY=$(usex numpy)
-DUSE_OPENCL=$(usex opencl)
-DUSE_OPENCV=$(usex opencv)
-DUSE_OPENMP=$(usex openmp)
-DUSE_ROCM=$(usex rocm)
-DUSE_SYSTEM_CPUINFO=ON
-DUSE_SYSTEM_PYBIND11=ON
-DUSE_UCC=OFF
-DUSE_VALGRIND=OFF
-DPYBIND11_PYTHON_VERSION="${EPYTHON#python}"
-DPYTHON_EXECUTABLE="${PYTHON}"
-DUSE_ITT=OFF
-DUSE_SYSTEM_PTHREADPOOL=ON
-DUSE_SYSTEM_FXDIV=ON
-DUSE_SYSTEM_FP16=ON
-DUSE_SYSTEM_GLOO=ON
-DUSE_SYSTEM_ONNX=ON
-DUSE_SYSTEM_SLEEF=ON
-DUSE_METAL=OFF
-Wno-dev
-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
)
if use mkl; then
mycmakeargs+=(-DBLAS=MKL)
elif use openblas; then
mycmakeargs+=(-DBLAS=OpenBLAS)
else
mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
fi
if use cuda; then
addpredict "/dev/nvidiactl" # bug 867706
addpredict "/dev/char"
mycmakeargs+=(
-DUSE_CUDNN=ON
-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
)
elif use rocm; then
export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
mycmakeargs+=(
-DUSE_NCCL=ON
-DUSE_SYSTEM_NCCL=ON
)
fi
if use onednn; then
mycmakeargs+=(
-DUSE_MKLDNN=ON
-DMKLDNN_FOUND=ON
-DMKLDNN_LIBRARIES=dnnl
-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
)
fi
cmake_src_configure
# do not rerun cmake and the build process in src_install
sed '/RERUN/,+1d' -i "${BUILD_DIR}"/build.ninja || die
}
src_install() {
cmake_src_install
insinto "/var/lib/${PN}"
doins "${BUILD_DIR}"/CMakeCache.txt
rm -rf python
mkdir -p python/torch/include || die
mv "${ED}"/usr/lib/python*/site-packages/caffe2 python/ || die
cp torch/version.py python/torch/ || die
python_domodule python/caffe2
python_domodule python/torch
ln -s ../../../../../include/torch \
"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
}
|
GitWeb