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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
|
{
stdenv,
lib,
fetchFromGitHub,
buildPythonPackage,
python,
config,
cudaSupport ? config.cudaSupport,
cudaPackages,
autoAddDriverRunpath,
effectiveMagma ?
if cudaSupport then
magma-cuda-static
else if rocmSupport then
magma-hip
else
magma,
magma,
magma-hip,
magma-cuda-static,
# Use the system NCCL as long as we're targeting CUDA on a supported platform.
useSystemNccl ? (cudaSupport && !cudaPackages.nccl.meta.unsupported || rocmSupport),
MPISupport ? false,
mpi,
buildDocs ? false,
# tests.cudaAvailable:
callPackage,
torchWithCuda,
# Native build inputs
cmake,
symlinkJoin,
which,
pybind11,
removeReferencesTo,
# Build inputs
numactl,
Accelerate,
CoreServices,
libobjc,
# Propagated build inputs
astunparse,
fsspec,
filelock,
jinja2,
networkx,
sympy,
numpy,
pyyaml,
cffi,
click,
typing-extensions,
# ROCm build and `torch.compile` requires `openai-triton`
tritonSupport ? (!stdenv.isDarwin),
openai-triton,
# Unit tests
hypothesis,
psutil,
# Disable MKLDNN on aarch64-darwin, it negatively impacts performance,
# this is also what official pytorch build does
mklDnnSupport ? !(stdenv.isDarwin && stdenv.isAarch64),
# virtual pkg that consistently instantiates blas across nixpkgs
# See https://github.com/NixOS/nixpkgs/pull/83888
blas,
# ninja (https://ninja-build.org) must be available to run C++ extensions tests,
ninja,
# dependencies for torch.utils.tensorboard
pillow,
six,
future,
tensorboard,
protobuf,
pythonOlder,
# ROCm dependencies
rocmSupport ? config.rocmSupport,
rocmPackages_5,
gpuTargets ? [ ],
}:
let
inherit (lib)
attrsets
lists
strings
trivial
;
inherit (cudaPackages) cudaFlags cudnn nccl;
rocmPackages = rocmPackages_5;
setBool = v: if v then "1" else "0";
# https://github.com/pytorch/pytorch/blob/v2.0.1/torch/utils/cpp_extension.py#L1744
supportedTorchCudaCapabilities =
let
real = [
"3.5"
"3.7"
"5.0"
"5.2"
"5.3"
"6.0"
"6.1"
"6.2"
"7.0"
"7.2"
"7.5"
"8.0"
"8.6"
"8.7"
"8.9"
"9.0"
];
ptx = lists.map (x: "${x}+PTX") real;
in
real ++ ptx;
# NOTE: The lists.subtractLists function is perhaps a bit unintuitive. It subtracts the elements
# of the first list *from* the second list. That means:
# lists.subtractLists a b = b - a
# For CUDA
supportedCudaCapabilities = lists.intersectLists cudaFlags.cudaCapabilities supportedTorchCudaCapabilities;
unsupportedCudaCapabilities = lists.subtractLists supportedCudaCapabilities cudaFlags.cudaCapabilities;
# Use trivial.warnIf to print a warning if any unsupported GPU targets are specified.
gpuArchWarner =
supported: unsupported:
trivial.throwIf (supported == [ ]) (
"No supported GPU targets specified. Requested GPU targets: "
+ strings.concatStringsSep ", " unsupported
) supported;
# Create the gpuTargetString.
gpuTargetString = strings.concatStringsSep ";" (
if gpuTargets != [ ] then
# If gpuTargets is specified, it always takes priority.
gpuTargets
else if cudaSupport then
gpuArchWarner supportedCudaCapabilities unsupportedCudaCapabilities
else if rocmSupport then
rocmPackages.clr.gpuTargets
else
throw "No GPU targets specified"
);
rocmtoolkit_joined = symlinkJoin {
name = "rocm-merged";
paths = with rocmPackages; [
rocm-core
clr
rccl
miopen
miopengemm
rocrand
rocblas
rocsparse
hipsparse
rocthrust
rocprim
hipcub
roctracer
rocfft
rocsolver
hipfft
hipsolver
hipblas
rocminfo
rocm-thunk
rocm-comgr
rocm-device-libs
rocm-runtime
clr.icd
hipify
];
# Fix `setuptools` not being found
postBuild = ''
rm -rf $out/nix-support
'';
};
brokenConditions = attrsets.filterAttrs (_: cond: cond) {
"CUDA and ROCm are mutually exclusive" = cudaSupport && rocmSupport;
"CUDA is not targeting Linux" = cudaSupport && !stdenv.isLinux;
"Unsupported CUDA version" =
cudaSupport
&& !(builtins.elem cudaPackages.cudaMajorVersion [
"11"
"12"
]);
"MPI cudatoolkit does not match cudaPackages.cudatoolkit" =
MPISupport && cudaSupport && (mpi.cudatoolkit != cudaPackages.cudatoolkit);
"Magma cudaPackages does not match cudaPackages" =
cudaSupport && (effectiveMagma.cudaPackages != cudaPackages);
"Rocm support is currently broken because `rocmPackages.hipblaslt` is unpackaged. (2024-06-09)" = rocmSupport;
};
in
buildPythonPackage rec {
pname = "torch";
# Don't forget to update torch-bin to the same version.
version = "2.3.1";
pyproject = true;
disabled = pythonOlder "3.8.0";
outputs = [
"out" # output standard python package
"dev" # output libtorch headers
"lib" # output libtorch libraries
"cxxdev" # propagated deps for the cmake consumers of torch
];
cudaPropagateToOutput = "cxxdev";
src = fetchFromGitHub {
owner = "pytorch";
repo = "pytorch";
rev = "refs/tags/v${version}";
fetchSubmodules = true;
hash = "sha256-vpgtOqzIDKgRuqdT8lB/g6j+oMIH1RPxdbjtlzZFjV8=";
};
patches =
lib.optionals cudaSupport [ ./fix-cmake-cuda-toolkit.patch ]
++ lib.optionals (stdenv.isDarwin && stdenv.isx86_64) [
# pthreadpool added support for Grand Central Dispatch in April
# 2020. However, this relies on functionality (DISPATCH_APPLY_AUTO)
# that is available starting with macOS 10.13. However, our current
# base is 10.12. Until we upgrade, we can fall back on the older
# pthread support.
./pthreadpool-disable-gcd.diff
]
++ lib.optionals stdenv.isLinux [
# Propagate CUPTI to Kineto by overriding the search path with environment variables.
# https://github.com/pytorch/pytorch/pull/108847
./pytorch-pr-108847.patch
];
postPatch =
lib.optionalString rocmSupport ''
# https://github.com/facebookincubator/gloo/pull/297
substituteInPlace third_party/gloo/cmake/Hipify.cmake \
--replace "\''${HIPIFY_COMMAND}" "python \''${HIPIFY_COMMAND}"
# Replace hard-coded rocm paths
substituteInPlace caffe2/CMakeLists.txt \
--replace "/opt/rocm" "${rocmtoolkit_joined}" \
--replace "hcc/include" "hip/include" \
--replace "rocblas/include" "include/rocblas" \
--replace "hipsparse/include" "include/hipsparse"
# Doesn't pick up the environment variable?
substituteInPlace third_party/kineto/libkineto/CMakeLists.txt \
--replace "\''$ENV{ROCM_SOURCE_DIR}" "${rocmtoolkit_joined}" \
--replace "/opt/rocm" "${rocmtoolkit_joined}"
# Strangely, this is never set in cmake
substituteInPlace cmake/public/LoadHIP.cmake \
--replace "set(ROCM_PATH \$ENV{ROCM_PATH})" \
"set(ROCM_PATH \$ENV{ROCM_PATH})''\nset(ROCM_VERSION ${lib.concatStrings (lib.intersperse "0" (lib.splitVersion rocmPackages.clr.version))})"
''
# Detection of NCCL version doesn't work particularly well when using the static binary.
+ lib.optionalString cudaSupport ''
substituteInPlace cmake/Modules/FindNCCL.cmake \
--replace \
'message(FATAL_ERROR "Found NCCL header version and library version' \
'message(WARNING "Found NCCL header version and library version'
''
# Remove PyTorch's FindCUDAToolkit.cmake and to use CMake's default.
# We do not remove the entirety of cmake/Modules_CUDA_fix because we need FindCUDNN.cmake.
+ lib.optionalString cudaSupport ''
rm cmake/Modules/FindCUDAToolkit.cmake
rm -rf cmake/Modules_CUDA_fix/{upstream,FindCUDA.cmake}
''
# error: no member named 'aligned_alloc' in the global namespace; did you mean simply 'aligned_alloc'
# This lib overrided aligned_alloc hence the error message. Tltr: his function is linkable but not in header.
+
lib.optionalString (stdenv.isDarwin && lib.versionOlder stdenv.hostPlatform.darwinSdkVersion "11.0")
''
substituteInPlace third_party/pocketfft/pocketfft_hdronly.h --replace-fail '#if (__cplusplus >= 201703L) && (!defined(__MINGW32__)) && (!defined(_MSC_VER))
inline void *aligned_alloc(size_t align, size_t size)' '#if 0
inline void *aligned_alloc(size_t align, size_t size)'
'';
# NOTE(@connorbaker): Though we do not disable Gloo or MPI when building with CUDA support, caution should be taken
# when using the different backends. Gloo's GPU support isn't great, and MPI and CUDA can't be used at the same time
# without extreme care to ensure they don't lock each other out of shared resources.
# For more, see https://github.com/open-mpi/ompi/issues/7733#issuecomment-629806195.
preConfigure =
lib.optionalString cudaSupport ''
export TORCH_CUDA_ARCH_LIST="${gpuTargetString}"
export CUPTI_INCLUDE_DIR=${lib.getDev cudaPackages.cuda_cupti}/include
export CUPTI_LIBRARY_DIR=${lib.getLib cudaPackages.cuda_cupti}/lib
''
+ lib.optionalString (cudaSupport && cudaPackages ? cudnn) ''
export CUDNN_INCLUDE_DIR=${lib.getLib cudnn}/include
export CUDNN_LIB_DIR=${cudnn.lib}/lib
''
+ lib.optionalString rocmSupport ''
export ROCM_PATH=${rocmtoolkit_joined}
export ROCM_SOURCE_DIR=${rocmtoolkit_joined}
export PYTORCH_ROCM_ARCH="${gpuTargetString}"
export CMAKE_CXX_FLAGS="-I${rocmtoolkit_joined}/include -I${rocmtoolkit_joined}/include/rocblas"
python tools/amd_build/build_amd.py
'';
# Use pytorch's custom configurations
dontUseCmakeConfigure = true;
# causes possible redefinition of _FORTIFY_SOURCE
hardeningDisable = [ "fortify3" ];
BUILD_NAMEDTENSOR = setBool true;
BUILD_DOCS = setBool buildDocs;
# We only do an imports check, so do not build tests either.
BUILD_TEST = setBool false;
# Unlike MKL, oneDNN (née MKLDNN) is FOSS, so we enable support for
# it by default. PyTorch currently uses its own vendored version
# of oneDNN through Intel iDeep.
USE_MKLDNN = setBool mklDnnSupport;
USE_MKLDNN_CBLAS = setBool mklDnnSupport;
# Avoid using pybind11 from git submodule
# Also avoids pytorch exporting the headers of pybind11
USE_SYSTEM_PYBIND11 = true;
# NB technical debt: building without NNPACK as workaround for missing `six`
USE_NNPACK = 0;
preBuild = ''
export MAX_JOBS=$NIX_BUILD_CORES
${python.pythonOnBuildForHost.interpreter} setup.py build --cmake-only
${cmake}/bin/cmake build
'';
preFixup = ''
function join_by { local IFS="$1"; shift; echo "$*"; }
function strip2 {
IFS=':'
read -ra RP <<< $(patchelf --print-rpath $1)
IFS=' '
RP_NEW=$(join_by : ''${RP[@]:2})
patchelf --set-rpath \$ORIGIN:''${RP_NEW} "$1"
}
for f in $(find ''${out} -name 'libcaffe2*.so')
do
strip2 $f
done
'';
# Override the (weirdly) wrong version set by default. See
# https://github.com/NixOS/nixpkgs/pull/52437#issuecomment-449718038
# https://github.com/pytorch/pytorch/blob/v1.0.0/setup.py#L267
PYTORCH_BUILD_VERSION = version;
PYTORCH_BUILD_NUMBER = 0;
# In-tree builds of NCCL are not supported.
# Use NCCL when cudaSupport is enabled and nccl is available.
USE_NCCL = setBool useSystemNccl;
USE_SYSTEM_NCCL = USE_NCCL;
USE_STATIC_NCCL = USE_NCCL;
# Suppress a weird warning in mkl-dnn, part of ideep in pytorch
# (upstream seems to have fixed this in the wrong place?)
# https://github.com/intel/mkl-dnn/commit/8134d346cdb7fe1695a2aa55771071d455fae0bc
# https://github.com/pytorch/pytorch/issues/22346
#
# Also of interest: pytorch ignores CXXFLAGS uses CFLAGS for both C and C++:
# https://github.com/pytorch/pytorch/blob/v1.11.0/setup.py#L17
env.NIX_CFLAGS_COMPILE = toString (
(
lib.optionals (blas.implementation == "mkl") [ "-Wno-error=array-bounds" ]
# Suppress gcc regression: avx512 math function raises uninitialized variable warning
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
# See also: Fails to compile with GCC 12.1.0 https://github.com/pytorch/pytorch/issues/77939
++ lib.optionals (stdenv.cc.isGNU && lib.versionAtLeast stdenv.cc.version "12.0.0") [
"-Wno-error=maybe-uninitialized"
"-Wno-error=uninitialized"
]
# Since pytorch 2.0:
# gcc-12.2.0/include/c++/12.2.0/bits/new_allocator.h:158:33: error: ‘void operator delete(void*, std::size_t)’
# ... called on pointer ‘<unknown>’ with nonzero offset [1, 9223372036854775800] [-Werror=free-nonheap-object]
++ lib.optionals (stdenv.cc.isGNU && lib.versions.major stdenv.cc.version == "12") [
"-Wno-error=free-nonheap-object"
]
# .../source/torch/csrc/autograd/generated/python_functions_0.cpp:85:3:
# error: cast from ... to ... converts to incompatible function type [-Werror,-Wcast-function-type-strict]
++ lib.optionals (stdenv.cc.isClang && lib.versionAtLeast stdenv.cc.version "16") [
"-Wno-error=cast-function-type-strict"
# Suppresses the most spammy warnings.
# This is mainly to fix https://github.com/NixOS/nixpkgs/issues/266895.
]
++ lib.optionals rocmSupport [
"-Wno-#warnings"
"-Wno-cpp"
"-Wno-unknown-warning-option"
"-Wno-ignored-attributes"
"-Wno-deprecated-declarations"
"-Wno-defaulted-function-deleted"
"-Wno-pass-failed"
]
++ [
"-Wno-unused-command-line-argument"
"-Wno-uninitialized"
"-Wno-array-bounds"
"-Wno-free-nonheap-object"
"-Wno-unused-result"
]
++ lib.optionals stdenv.cc.isGNU [
"-Wno-maybe-uninitialized"
"-Wno-stringop-overflow"
]
)
);
nativeBuildInputs =
[
cmake
which
ninja
pybind11
removeReferencesTo
]
++ lib.optionals cudaSupport (
with cudaPackages;
[
autoAddDriverRunpath
cuda_nvcc
]
)
++ lib.optionals rocmSupport [ rocmtoolkit_joined ];
buildInputs =
[
blas
blas.provider
]
++ lib.optionals cudaSupport (
with cudaPackages;
[
cuda_cccl # <thrust/*>
cuda_cudart # cuda_runtime.h and libraries
cuda_cupti # For kineto
cuda_nvcc # crt/host_config.h; even though we include this in nativeBuildinputs, it's needed here too
cuda_nvml_dev # <nvml.h>
cuda_nvrtc
cuda_nvtx # -llibNVToolsExt
libcublas
libcufft
libcurand
libcusolver
libcusparse
]
++ lists.optionals (cudaPackages ? cudnn) [
cudnn
]
++ lists.optionals useSystemNccl [
# Some platforms do not support NCCL (i.e., Jetson)
nccl # Provides nccl.h AND a static copy of NCCL!
]
++ lists.optionals (strings.versionOlder cudaVersion "11.8") [
cuda_nvprof # <cuda_profiler_api.h>
]
++ lists.optionals (strings.versionAtLeast cudaVersion "11.8") [
cuda_profiler_api # <cuda_profiler_api.h>
]
)
++ lib.optionals rocmSupport [ rocmPackages.llvm.openmp ]
++ lib.optionals (cudaSupport || rocmSupport) [ effectiveMagma ]
++ lib.optionals stdenv.isLinux [ numactl ]
++ lib.optionals stdenv.isDarwin [
Accelerate
CoreServices
libobjc
]
++ lib.optionals tritonSupport [ openai-triton ]
++ lib.optionals MPISupport [ mpi ]
++ lib.optionals rocmSupport [ rocmtoolkit_joined ];
dependencies = [
astunparse
cffi
click
numpy
pyyaml
# From install_requires:
fsspec
filelock
typing-extensions
sympy
networkx
jinja2
# the following are required for tensorboard support
pillow
six
future
tensorboard
protobuf
# torch/csrc requires `pybind11` at runtime
pybind11
] ++ lib.optionals tritonSupport [ openai-triton ];
propagatedCxxBuildInputs =
[ ] ++ lib.optionals MPISupport [ mpi ] ++ lib.optionals rocmSupport [ rocmtoolkit_joined ];
# Tests take a long time and may be flaky, so just sanity-check imports
doCheck = false;
pythonImportsCheck = [ "torch" ];
nativeCheckInputs = [
hypothesis
ninja
psutil
];
checkPhase =
with lib.versions;
with lib.strings;
concatStringsSep " " [
"runHook preCheck"
"${python.interpreter} test/run_test.py"
"--exclude"
(concatStringsSep " " [
"utils" # utils requires git, which is not allowed in the check phase
# "dataloader" # psutils correctly finds and triggers multiprocessing, but is too sandboxed to run -- resulting in numerous errors
# ^^^^^^^^^^^^ NOTE: while test_dataloader does return errors, these are acceptable errors and do not interfere with the build
# tensorboard has acceptable failures for pytorch 1.3.x due to dependencies on tensorboard-plugins
(optionalString (majorMinor version == "1.3") "tensorboard")
])
"runHook postCheck"
];
pythonRemoveDeps = [
# In our dist-info the name is just "triton"
"pytorch-triton-rocm"
];
postInstall =
''
find "$out/${python.sitePackages}/torch/include" "$out/${python.sitePackages}/torch/lib" -type f -exec remove-references-to -t ${stdenv.cc} '{}' +
mkdir $dev
cp -r $out/${python.sitePackages}/torch/include $dev/include
cp -r $out/${python.sitePackages}/torch/share $dev/share
# Fix up library paths for split outputs
substituteInPlace \
$dev/share/cmake/Torch/TorchConfig.cmake \
--replace \''${TORCH_INSTALL_PREFIX}/lib "$lib/lib"
substituteInPlace \
$dev/share/cmake/Caffe2/Caffe2Targets-release.cmake \
--replace \''${_IMPORT_PREFIX}/lib "$lib/lib"
mkdir $lib
mv $out/${python.sitePackages}/torch/lib $lib/lib
ln -s $lib/lib $out/${python.sitePackages}/torch/lib
''
+ lib.optionalString rocmSupport ''
substituteInPlace $dev/share/cmake/Tensorpipe/TensorpipeTargets-release.cmake \
--replace "\''${_IMPORT_PREFIX}/lib64" "$lib/lib"
substituteInPlace $dev/share/cmake/ATen/ATenConfig.cmake \
--replace "/build/source/torch/include" "$dev/include"
'';
postFixup =
''
mkdir -p "$cxxdev/nix-support"
printWords "''${propagatedCxxBuildInputs[@]}" >> "$cxxdev/nix-support/propagated-build-inputs"
''
+ lib.optionalString stdenv.isDarwin ''
for f in $(ls $lib/lib/*.dylib); do
install_name_tool -id $lib/lib/$(basename $f) $f || true
done
install_name_tool -change @rpath/libshm.dylib $lib/lib/libshm.dylib $lib/lib/libtorch_python.dylib
install_name_tool -change @rpath/libtorch.dylib $lib/lib/libtorch.dylib $lib/lib/libtorch_python.dylib
install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libtorch_python.dylib
install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libtorch.dylib
install_name_tool -change @rpath/libtorch.dylib $lib/lib/libtorch.dylib $lib/lib/libshm.dylib
install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libshm.dylib
'';
# See https://github.com/NixOS/nixpkgs/issues/296179
#
# This is a quick hack to add `libnvrtc` to the runpath so that torch can find
# it when it is needed at runtime.
extraRunpaths = lib.optionals cudaSupport [ "${lib.getLib cudaPackages.cuda_nvrtc}/lib" ];
postPhases = lib.optionals stdenv.isLinux [ "postPatchelfPhase" ];
postPatchelfPhase = ''
while IFS= read -r -d $'\0' elf ; do
for extra in $extraRunpaths ; do
echo patchelf "$elf" --add-rpath "$extra" >&2
patchelf "$elf" --add-rpath "$extra"
done
done < <(
find "''${!outputLib}" "$out" -type f -iname '*.so' -print0
)
'';
# Builds in 2+h with 2 cores, and ~15m with a big-parallel builder.
requiredSystemFeatures = [ "big-parallel" ];
passthru = {
inherit
cudaSupport
cudaPackages
rocmSupport
rocmPackages
;
cudaCapabilities = if cudaSupport then supportedCudaCapabilities else [ ];
# At least for 1.10.2 `torch.fft` is unavailable unless BLAS provider is MKL. This attribute allows for easy detection of its availability.
blasProvider = blas.provider;
# To help debug when a package is broken due to CUDA support
inherit brokenConditions;
tests = callPackage ./tests.nix { };
};
meta = {
changelog = "https://github.com/pytorch/pytorch/releases/tag/v${version}";
# keep PyTorch in the description so the package can be found under that name on search.nixos.org
description = "PyTorch: Tensors and Dynamic neural networks in Python with strong GPU acceleration";
homepage = "https://pytorch.org/";
license = lib.licenses.bsd3;
maintainers = with lib.maintainers; [
teh
thoughtpolice
tscholak
]; # tscholak esp. for darwin-related builds
platforms = with lib.platforms; linux ++ lib.optionals (!cudaSupport && !rocmSupport) darwin;
broken = builtins.any trivial.id (builtins.attrValues brokenConditions);
};
}
|