gentoo-mirror/gentoo
1
0
Fork 0
mirror of https://github.com/gentoo-mirror/gentoo.git synced 2026-08-06 00:48:18 -07:00
Code Issues Projects Releases Wiki Activity

Merge updates from master

Browse Source
This commit is contained in:
Repository mirror & CI
2026-08-05 15:45:47 +00:00
parent ebbdd9134c 86ceb0ae0c
commit 0e7e229278
9 changed files with 0 additions and 820 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
sci-ml/caffe2/Manifest
View File

@@ -1,4 +1,3 @@
DIST composable_kernel-7fe50dc3.tar.gz 5380728 BLAKE2B c89c346d8e2d7a93a9cf26409e477fcdd25c43bc3f99d904c3bfe1bc282c6844ef2f2c80aceabe3bf4494db3457285384d5de5a22281aa426ba7479af82b0caf SHA512 a62f92e2dd7da944bd34bab6cf3bf624f630dc316d29c755e9fd523343c3f7648b0b7e0c9a0c8f5e9654477599ae8be9dac687d4054b0390f064ac2e40fc1cd3
DIST flash-attention-2.7.4.gh.tar.gz 5841323 BLAKE2B 432999d763f2b3d732580ddfea5d3e01370351db0656546259a5e500a07516dd03c98828bfb55855dabe4adc651033b5d97ea4725ca46158b9970f0fbc662710 SHA512 05a4afb09e666f7404d6a3f8b5256e7bed6eba60a6f1bde2b7dbb96d318975f0b458c2521c7a38d88e97b6e4c27f29077cf787849daf82586e33f43a3d9a84b3
DIST pytorch-2.11.0.tar.gz 63504636 BLAKE2B d7227bd2d5475acf696bb48a3ca8cd8b7b6cd75fb99e64ebc71434e23817f94dea5164f48a176dcad1acd12186c33b02452582c1b41d6265b89cf54b29630670 SHA512 a5b27efdbce0506b49808e9dc5fbf1a56a22b1651b60236570823e1fe2c7395a0de9d3d7a01a568a201d593c52e7aa402e5d7645a54cb5932d5866e374ebf945
DIST pytorch-2.12.0.tar.gz 64740318 BLAKE2B e5e46a93422454da420563d0a8057c87d1ed4b27e0bd9b3b30e46b5e40cfbb489072144ac00267af2f085432d55e12adf750daf15ea165a1252449b00dca1ff9 SHA512 f5ab0f6933d88271f772b416f8c9b3b0d3e1ffaf8d00838b455206266b40d7c805e34003d84b01cddc7f1ad917dd72d6f79a05063b0962cbf223d9042cff3206

420
sci-ml/caffe2/caffe2-2.11.0-r3.ebuild
View File

@@ -1,420 +0,0 @@
# Copyright 2022-2026 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2
EAPI=8
PYTHON_COMPAT=( python3_{11..14} )
ROCM_VERSION=6.1
inherit python-single-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
MYPN=pytorch
MYP=${MYPN}-${PV}
# caffe2-2.9.0 depends on future version of composable kernel
# TODO: replace it with DEPEND in the future
CK_COMMIT=7fe50dc3da2069d6645d9deb8c017a876472a977
CK_P=composable_kernel-${CK_COMMIT:0:8}
FLASH_PV=2.7.4
FLASH_PN=flash-attention
FLASH_P=${FLASH_PN}-${FLASH_PV}
FLASH_ATT_URI="https://github.com/Dao-AILab/${FLASH_PN}/archive/refs/tags/v${FLASH_PV}.tar.gz -> ${FLASH_P}.gh.tar.gz"
AOTRITON_PV=0.9.2b
AOTRITON_PN=aotriton
AOTRITON_P=${AOTRITON_PN}-${AOTRITON_PV}
AOTRITON_tar=${AOTRITON_P}-manylinux_2_28_x86_64-rocm6.3-shared.tar.gz
DESCRIPTION="A deep learning framework"
HOMEPAGE="https://pytorch.org/"
SRC_URI="
https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz -> ${MYP}.tar.gz
rocm? (
https://github.com/ROCm/composable_kernel/archive/${CK_COMMIT}.tar.gz
-> ${CK_P}.tar.gz
)
cuda? (
flash? ( ${FLASH_ATT_URI} )
memefficient? ( ${FLASH_ATT_URI} )
)
"
S="${WORKDIR}"/${MYP}
LICENSE="BSD"
SLOT="0"
KEYWORDS="~amd64 ~arm64"
IUSE="cuda cusparselt distributed fbgemm flash gloo kineto memefficient
mimalloc mkl mpi nccl nnpack +numpy onednn openblas opencl openmp qnnpack
rocm xnnpack"
RESTRICT="test"
REQUIRED_USE="
${PYTHON_REQUIRED_USE}
mpi? ( distributed )
gloo? ( distributed )
?? ( cuda rocm )
rocm? (
|| ( ${ROCM_REQUIRED_USE} )
memefficient? ( flash )
)
cusparselt? ( || ( cuda rocm ) )
flash? ( || ( cuda rocm ) )
memefficient? ( || ( cuda rocm ) )
nccl? ( rocm )
"
RDEPEND="
${PYTHON_DEPS}
dev-cpp/abseil-cpp:=
dev-cpp/gflags:=
>=dev-cpp/glog-0.5.0:=
>=dev-libs/cpuinfo-2025.11.14
dev-libs/libfmt:=
dev-libs/protobuf:=
dev-libs/sleef
sci-ml/onnx
virtual/lapack
cuda? (
dev-libs/cudnn
>=sci-ml/cudnn-frontend-1.12.0:=
>=dev-util/nvidia-cuda-toolkit-12.9:=[profiler]
cusparselt? ( dev-libs/cusparselt )
)
fbgemm? ( >=sci-ml/FBGEMM-1.4 )
gloo? ( >=sci-ml/gloo-2025.06.04[cuda?,rocm?] )
kineto? ( ~sci-ml/kineto-0.4.0_p20260323 )
mimalloc? ( dev-libs/mimalloc )
mpi? ( virtual/mpi )
nnpack? (
sci-ml/NNPACK
dev-libs/pthreadpool
)
numpy? ( $(python_gen_cond_dep '
dev-python/numpy[${PYTHON_USEDEP}]
') )
onednn? ( sci-ml/oneDNN )
opencl? ( virtual/opencl )
qnnpack? (
!sci-libs/QNNPACK
sci-ml/gemmlowp
dev-libs/pthreadpool
)
rocm? (
nccl? ( >=dev-libs/rccl-6.3:= <dev-libs/rccl-7.3:= )
>=dev-util/hip-6.3:= <dev-util/hip-7.3:=
>=dev-util/roctracer-6.3:= <dev-util/roctracer-7.3:=
>=sci-libs/hipBLAS-6.3:= <sci-libs/hipBLAS-7.3:=[rocsolver(+)]
>=sci-libs/hipBLASLt-6.3:= <sci-libs/hipBLASLt-7.3:=
>=sci-libs/hipFFT-6.3:= <sci-libs/hipFFT-7.3:=
>=sci-libs/hipRAND-6.3:= <sci-libs/hipRAND-7.3:=
>=sci-libs/hipSOLVER-6.3:= <sci-libs/hipSOLVER-7.3:=
>=sci-libs/hipSPARSE-6.3:= <sci-libs/hipSPARSE-7.3:=
>=sci-libs/miopen-6.3:= <sci-libs/miopen-7.3:=
>=sci-libs/rocBLAS-6.3:= <sci-libs/rocBLAS-7.3:=
>=sci-libs/rocRAND-6.3:= <sci-libs/rocRAND-7.3:=
>=sci-libs/rocSOLVER-6.3:= <sci-libs/rocSOLVER-7.3:=
memefficient? ( =sci-libs/aotriton-bin-0.11*:= )
distributed? ( >=dev-util/rocm-smi-6.3:= <dev-util/rocm-smi-7.3:= )
cusparselt? ( >=sci-libs/hipsparselt-6.3:= <sci-libs/hipsparselt-7.3:= )
)
distributed? (
!rocm? ( sci-ml/tensorpipe[cuda?] )
dev-cpp/cpp-httplib:=
)
xnnpack? (
>=sci-ml/XNNPACK-2024.11
dev-libs/pthreadpool
)
mkl? ( sci-libs/mkl )
openblas? ( sci-libs/openblas )
"
DEPEND="
${RDEPEND}
dev-cpp/nlohmann_json
dev-libs/flatbuffers
dev-libs/FXdiv
dev-libs/pocketfft
dev-libs/psimd
sci-ml/FP16
$(python_gen_cond_dep '
<dev-python/pybind11-3.0.2[${PYTHON_USEDEP}]
dev-python/pyyaml[${PYTHON_USEDEP}]
dev-python/typing-extensions[${PYTHON_USEDEP}]
')
cuda? ( >=dev-libs/cutlass-3.9.2[tools(+)] )
onednn? ( sci-ml/ideep )
rocm? (
>=sci-libs/hipCUB-6.3:= <sci-libs/hipCUB-7.3:=
>=sci-libs/rocPRIM-6.3:= <sci-libs/rocPRIM-7.3:=
>=sci-libs/rocThrust-6.3:= <sci-libs/rocThrust-7.3:=
)
qnnpack? ( dev-libs/clog )
"
PATCHES=(
"${FILESDIR}"/${PN}-2.5.1-unbundle_fmt.patch
"${FILESDIR}"/${PN}-2.5.1-unbundle_kineto.patch
"${FILESDIR}"/${PN}-2.8.0-unbundle_pocketfft.patch
"${FILESDIR}"/${PN}-2.5.1-cudnn_include_fix.patch
"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
"${FILESDIR}"/${PN}-2.5.1-glog-0.6.0.patch
"${FILESDIR}"/${P}-rocm-fix-std-cpp17.patch
"${FILESDIR}"/${PN}-2.7.0-glog-0.7.1.patch
"${FILESDIR}"/${PN}-2.7.1-aotriton-fixes.patch
"${FILESDIR}"/${PN}-2.8.0-rocm-minus-flash.patch
"${FILESDIR}"/${PN}-2.9.0-cmake.patch
"${FILESDIR}"/${PN}-2.9.0-rocm-distributed-link.patch
"${FILESDIR}"/${PN}-2.9.1-torch_cpu.patch
"${FILESDIR}"/${PN}-2.10.0-gentoo.patch
"${FILESDIR}"/${P}-mimalloc.patch
"${FILESDIR}"/${P}-removekineto-pr178960.patch
)
src_prepare() {
if use cuda && ( use flash || use memefficient ); then
mv "${WORKDIR}"/${FLASH_P}/* third_party/${FLASH_PN}/ || die
fi
filter-lto #bug 862672
# Unbundle fmt
sed -i \
-e 's|::fmt-header-only||' \
c10/CMakeLists.txt \
cmake/Dependencies.cmake \
torch/CMakeLists.txt \
|| die
# tensorpipe is in system, not a build target of caffe2
sed -e '/target_compile_options_if_supported(tensorpipe/d' -i cmake/Dependencies.cmake || die
# Drop third_party from CMake tree
sed -i \
-e '/add_subdirectory.*third_party/d' \
CMakeLists.txt \
cmake/Dependencies.cmake \
cmake/ProtoBuf.cmake \
aten/src/ATen/CMakeLists.txt \
|| die
# Change libc10* path
sed -i \
-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
c10/cuda/CMakeLists.txt \
c10/CMakeLists.txt \
c10/hip/CMakeLists.txt \
|| die
# Change libaotriton path
sed -i \
-e "s|}/lib|}/\${CMAKE_INSTALL_LIBDIR}|g" \
-e "/set(__AOTRITON_LIB/s|lib/|\${CMAKE_INSTALL_LIBDIR}/|g" \
cmake/External/aotriton.cmake \
|| die
# Noisy warnings from Logging.h
sed -i 's/-Wextra-semi//' cmake/public/utils.cmake || die
cmake_src_prepare
pushd torch/csrc/jit/serialization > /dev/null || die
flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
popd > /dev/null || die
# prefixify the hardcoded paths, after all patches are applied
hprefixify \
aten/CMakeLists.txt \
caffe2/CMakeLists.txt \
cmake/Metal.cmake \
cmake/Modules/*.cmake \
cmake/Modules_CUDA_fix/FindCUDNN.cmake \
cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
cmake/public/LoadHIP.cmake \
cmake/public/cuda.cmake \
cmake/Dependencies.cmake \
torch/CMakeLists.txt \
CMakeLists.txt
if use rocm; then
sed -e "s:/opt/rocm:/usr:" \
-e "s:lib/cmake:$(get_libdir)/cmake:g" \
-i cmake/public/LoadHIP.cmake || die
# TODO: delete, when caffe2 depends on systemwide composable_kernel
sed -e "s:third_party/composable_kernel:../composable_kernel-${CK_COMMIT}:g" \
-i aten/src/ATen/CMakeLists.txt || die
# Bug 959808: fix for gfx101x targets
pushd "${WORKDIR}/composable_kernel-${CK_COMMIT}" > /dev/null || die
eapply "${FILESDIR}"/composable-kernel-7fe50dc-expand-isa.patch
popd > /dev/null || die
if tc-is-clang; then
# Systemwide gcc (for absl and at::TensorBase) + hipcc (llvm>=18) need abi-compat=17.
# But systemwide clang>=18 + hipcc (>=llvm-18) need opposite!
# See also: https://github.com/llvm/llvm-project/issues/102443#issuecomment-2329726287
sed -e '/-fclang-abi-compat=17/d' -i cmake/Dependencies.cmake || die
fi
# Workaround for libc++ issue https://github.com/llvm/llvm-project/issues/100802
sed -e 's/std::memcpy/memcpy/g' -i torch/headeronly/util/Half.h || die
ebegin "HIPifying cuda sources"
FBCODE_BUILD_TOOL="buck" ${EPYTHON} tools/amd_build/build_amd.py || die
eend $?
fi
}
src_configure() {
if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
ewarn "These may not be optimal for your GPU."
ewarn ""
ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
ewarn ""
ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
fi
local mycmakeargs=(
-DBUILD_CUSTOM_PROTOBUF=OFF
-DBUILD_TEST=OFF
-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
-DPython_EXECUTABLE="${PYTHON}"
-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
-DUSE_CCACHE=OFF
-DUSE_CUDA=$(usex cuda)
-DUSE_DISTRIBUTED=$(usex distributed)
-DUSE_FBGEMM=$(usex fbgemm)
-DUSE_FLASH_ATTENTION=$(usex flash)
-DUSE_GFLAGS=ON
-DUSE_GLOG=ON
-DUSE_GLOO=$(usex gloo)
-DUSE_ITT=OFF
-DUSE_KINETO=$(usex kineto)
-DUSE_KLEIDIAI=OFF # TODO
-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
-DUSE_MIMALLOC=$(usex mimalloc)
-DUSE_MKLDNN=$(usex onednn)
-DUSE_MPI=$(usex mpi)
-DUSE_NCCL=OFF
-DUSE_NNPACK=$(usex nnpack)
-DUSE_NUMA=OFF
-DUSE_NUMPY=$(usex numpy)
-DUSE_OPENCL=$(usex opencl)
-DUSE_OPENMP=$(usex openmp)
-DUSE_PYTORCH_QNNPACK=$(usex qnnpack)
-DUSE_PYTORCH_METAL=OFF
-DUSE_ROCM=$(usex rocm)
-DUSE_SYSTEM_CPUINFO=ON
-DUSE_SYSTEM_EIGEN_INSTALL=ON
-DUSE_SYSTEM_FP16=ON
-DUSE_SYSTEM_FXDIV=ON
-DUSE_SYSTEM_GLOO=ON
-DUSE_SYSTEM_NVTX=ON
-DUSE_SYSTEM_ONNX=ON
-DUSE_SYSTEM_PSIMD=ON
-DUSE_SYSTEM_PTHREADPOOL=ON
-DUSE_SYSTEM_PYBIND11=ON
-DUSE_SYSTEM_SLEEF=ON
-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
-DUSE_TENSORPIPE=$(usex distributed $(usex !rocm))
-DUSE_UCC=OFF
-DUSE_VALGRIND=OFF
-DUSE_XNNPACK=$(usex xnnpack)
-DUSE_XPU=OFF
-Wno-dev
)
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
# bug 867706 926116
cuda_add_sandbox
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 \")"
-DUSE_CUSPARSELT=$(usex cusparselt)
)
[[ -v CUDACXX ]] && export PYTORCH_NVCC="${CUDACXX}"
if use flash; then
export FLASH_ATTENTION_FORCE_BUILD="TRUE"
export FLASH_ATTN_CUDA_ARCHS="${CUDAARCHS:-${TORCH_CUDA_ARCH_LIST:-3.5 7.0}}"
fi
elif use rocm; then
export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
if use memefficient; then
export AOTRITON_INSTALLED_PREFIX="${ESYSROOT}/usr"
fi
mycmakeargs+=(
-DUSE_NCCL=$(usex nccl)
-DUSE_SYSTEM_NCCL=ON
-DCMAKE_REQUIRE_FIND_PACKAGE_HIP=ON
-DCMAKE_DISABLE_FIND_PACKAGE_hipsparselt=$(usex !cusparselt) # disable automagic
-DUSE_ROCM_CK_SDPA=OFF # requires flash + aiter, works only on gfx90a/gfx942/gfx950
)
# ROCm libraries produce too much warnings
append-cxxflags -Wno-deprecated-declarations -Wno-unused-result -Wno-unused-value
fi
if use onednn; then
mycmakeargs+=(
-DMKLDNN_FOUND=ON
-DMKLDNN_LIBRARIES=dnnl
-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
)
fi
cmake_src_configure
}
src_compile() {
PYTORCH_BUILD_VERSION=${PV} \
PYTORCH_BUILD_NUMBER=0 \
cmake_src_compile
}
python_install() {
python_domodule python/torch
mkdir "${D}"$(python_get_sitedir)/torch/bin || die
mkdir "${D}"$(python_get_sitedir)/torch/lib || die
mkdir "${D}"$(python_get_sitedir)/torch/include || die
ln -s ../../../../../include/torch \
"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
ln -s ../../../../../bin/torch_shm_manager \
"${D}"/$(python_get_sitedir)/torch/bin/torch_shm_manager || die
ln -s ../../../../../$(get_libdir)/libtorch_global_deps.so \
"${D}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
}
src_install() {
cmake_src_install
# Used by pytorch ebuild
insinto "/var/lib/${PN}"
doins "${BUILD_DIR}"/CMakeCache.txt
rm -rf python
mkdir -p python/torch || die
cp torch/version.py python/torch/ || die
python_install
}

170
sci-ml/caffe2/files/caffe2-2.11.0-removekineto-pr178960.patch
View File

@@ -1,170 +0,0 @@
diff --git a/torch/csrc/autograd/init.cpp b/torch/csrc/autograd/init.cpp
index 28c1a43..70d8f7b 100644
--- a/torch/csrc/autograd/init.cpp
+++ b/torch/csrc/autograd/init.cpp
@@ -301,12 +301,7 @@ PyObject* THPAutograd_initExtension(PyObject* _unused, PyObject* unused) {
.def("privateuse1_elapsed_us", &KinetoEvent::privateuse1ElapsedUs)
.def(
"is_user_annotation",
- [](const KinetoEvent& e) {
- return e.activityType() ==
- (uint8_t)libkineto::ActivityType::USER_ANNOTATION ||
- e.activityType() ==
- (uint8_t)libkineto::ActivityType::GPU_USER_ANNOTATION;
- })
+ [](const KinetoEvent& e) { return e.isUserAnnotation(); })
.def("nbytes", [](const KinetoEvent& e) { return e.nBytes(); })
// whether the event is hidden
.def(
diff --git a/torch/csrc/autograd/profiler_kineto.cpp b/torch/csrc/autograd/profiler_kineto.cpp
index dd4598669827e..85de1c3af923f 100644
--- a/torch/csrc/autograd/profiler_kineto.cpp
+++ b/torch/csrc/autograd/profiler_kineto.cpp
@@ -1081,6 +1081,13 @@ int64_t KinetoEvent::privateuse1ElapsedUs() const {
return -1;
}
+bool KinetoEvent::isUserAnnotation() const {
+ constexpr uint8_t kUserAnnotation = 1;
+ constexpr uint8_t kGpuUserAnnotation = 2;
+ const auto type = activityType();
+ return type == kUserAnnotation || type == kGpuUserAnnotation;
+}
+
void KinetoEvent::getPerfEventCounters(std::vector<uint64_t>& in) const {
return result_->visit(c10::overloaded(
[&in](const ExtraFields<EventType::TorchOp>& e) -> void {
diff --git a/torch/csrc/autograd/profiler_kineto.h b/torch/csrc/autograd/profiler_kineto.h
index 777b8a5851ed5..c3135c61245d4 100644
--- a/torch/csrc/autograd/profiler_kineto.h
+++ b/torch/csrc/autograd/profiler_kineto.h
@@ -67,6 +67,7 @@ struct TORCH_API KinetoEvent {
bool isPythonFunction() const;
int64_t cudaElapsedUs() const;
int64_t privateuse1ElapsedUs() const;
+ bool isUserAnnotation() const;
void getPerfEventCounters(torch::profiler::perf_counters_t& /*in*/) const;
extra_meta_t extraMeta() const;
std::string metadataJson() const;
diff --git a/torch/csrc/profiler/collection.cpp b/torch/csrc/profiler/collection.cpp
index eb29cb7382127..8ee3a9e7707a8 100644
--- a/torch/csrc/profiler/collection.cpp
+++ b/torch/csrc/profiler/collection.cpp
@@ -572,11 +572,13 @@ std::string toString(const ExtraFields<EventType::PyCall>& e) {
e.callsite_.funcname_.str());
}
+#ifdef USE_KINETO
auto scopeToType(at::RecordScope scope) {
return scope == at::RecordScope::USER_SCOPE
? libkineto::ActivityType::USER_ANNOTATION
: libkineto::ActivityType::CPU_OP;
}
+#endif
int64_t torchOpEndNS(
const ExtraFields<EventType::TorchOp>& e,
@@ -625,6 +627,7 @@ std::string Result::overload_name() const {
[](const auto& e) -> std::string { return ""; }));
}
+#ifdef USE_KINETO
libkineto::ActivityType Result::kinetoType() const {
return visit(c10::overloaded(
ATTRIBUTE(TorchOp, scopeToType(e.scope_)),
@@ -637,6 +640,11 @@ libkineto::ActivityType Result::kinetoType() const {
ATTRIBUTE(PythonGC, libkineto::ActivityType::PYTHON_FUNCTION),
ATTRIBUTE(Kineto, e.activity_type_)));
}
+#else
+libkineto::ActivityType Result::kinetoType() const {
+ return libkineto::ActivityType::NONE;
+}
+#endif
uint64_t Result::correlationID() const {
return visit(c10::overloaded(
diff --git a/torch/csrc/profiler/kineto_shim.cpp b/torch/csrc/profiler/kineto_shim.cpp
index fa232e1a01016..0ddf62997a07e 100644
--- a/torch/csrc/profiler/kineto_shim.cpp
+++ b/torch/csrc/profiler/kineto_shim.cpp
@@ -138,6 +138,7 @@ TraceWrapper::TraceWrapper(const int64_t start_time, const std::string& name)
}
#endif // USE_KINETO
+#ifdef USE_KINETO
activity_t* TraceWrapper::addCPUActivity(
const std::string& name,
const libkineto::ActivityType type,
@@ -145,7 +146,6 @@ activity_t* TraceWrapper::addCPUActivity(
const uint64_t correlation_id,
const int64_t start_time,
const int64_t end_time) {
-#ifdef USE_KINETO
TORCH_CHECK((bool)(*this), "Cannot add event to non-existent trace.");
cpu_trace_->emplace_activity(cpu_trace_->span, type, name);
auto& act = libkineto::CpuTraceBuffer::toRef(cpu_trace_->activities.back());
@@ -157,10 +157,18 @@ activity_t* TraceWrapper::addCPUActivity(
act.endTime = end_time;
}
return cpu_trace_->activities.back().get();
+}
#else
+activity_t* TraceWrapper::addCPUActivity(
+ const std::string& name,
+ const libkineto::ActivityType type,
+ const DeviceAndResource device_and_resource,
+ const uint64_t correlation_id,
+ const int64_t start_time,
+ const int64_t end_time) {
return nullptr;
-#endif // USE_KINETO
}
+#endif // USE_KINETO
void TraceWrapper::transferCpuTrace(int64_t end_time) {
#ifdef USE_KINETO
@@ -473,6 +481,7 @@ void logInvariantViolation(
namespace autograd::profiler {
c10::DeviceType deviceTypeFromActivity(libkineto::ActivityType activity_type) {
+#ifdef USE_KINETO
// PrivateUse1 kineto backend reuse some ActivityTypes,
// If PrivateUse1 backend is enabled, this should return
// c10::DeviceType::PrivateUse1.
@@ -524,6 +533,9 @@ c10::DeviceType deviceTypeFromActivity(libkineto::ActivityType activity_type) {
return c10::DeviceType::CPU;
}
}
+#else
+ return c10::DeviceType::CPU;
+#endif // USE_KINETO
}
void addMetadataJson(const std::string& key, const std::string& value) {
diff --git a/torch/csrc/profiler/kineto_shim.h b/torch/csrc/profiler/kineto_shim.h
index 4f9bdc6770507..ddc00c03fa9cc 100644
--- a/torch/csrc/profiler/kineto_shim.h
+++ b/torch/csrc/profiler/kineto_shim.h
@@ -12,7 +12,21 @@
#undef USE_KINETO
#endif
+#ifdef USE_KINETO
#include <ActivityType.h>
+#else
+namespace libkineto {
+enum class ActivityType : uint8_t {
+ CPU_OP = 0,
+ USER_ANNOTATION,
+ GPU_USER_ANNOTATION,
+ NONE = CPU_OP,
+};
+inline const char* toString(ActivityType) {
+ return "CPU_OP";
+}
+} // namespace libkineto
+#endif
#include <torch/csrc/Export.h>
#include <torch/csrc/profiler/api.h>

50
sci-ml/caffe2/files/caffe2-2.11.0-rocm-fix-std-cpp17.patch
View File

@@ -1,50 +0,0 @@
Fix for CXX=clang USE=rocm error: invalid argument '-std=c++17' not allowed with 'C'
https://github.com/pytorch/pytorch/issues/103222
--- a/c10/hip/CMakeLists.txt
+++ b/c10/hip/CMakeLists.txt
@@ -37,6 +37,7 @@ if(NOT BUILD_LIBTORCHLESS)
# Propagate HIP_CXX_FLAGS that were set from Dependencies.cmake
target_compile_options(c10_hip PRIVATE ${HIP_CXX_FLAGS})
+ set_target_properties(c10_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
# caffe2_hip adds a bunch of dependencies like rocsparse, but c10/hip is supposed to be
# minimal. I'm not sure if we need hip_hcc or not; for now leave it out
--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
@@ -1774,6 +1774,7 @@ if(USE_ROCM)
# Since PyTorch files contain HIP headers, these flags are required for the necessary definitions to be added.
target_compile_options(torch_hip PUBLIC ${HIP_CXX_FLAGS}) # experiment
+ set_target_properties(torch_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
target_link_libraries(torch_hip PUBLIC c10_hip)
@@ -2003,6 +2004,7 @@ if(BUILD_TEST)
target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE} ${Caffe2_HIP_INCLUDE})
target_compile_options(${test_name} PRIVATE ${HIP_CXX_FLAGS})
+ set_target_properties(${test_name} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
if(INSTALL_TEST)
set_target_properties(${test_name} PROPERTIES INSTALL_RPATH "${CMAKE_INSTALL_RPATH}:${_rpath_portable_origin}/../lib")
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -1014,7 +1014,6 @@
list(APPEND HIP_CXX_FLAGS -DUSE_ROCM_CK_GEMM)
endif()
list(APPEND HIP_HIPCC_FLAGS --offload-compress)
- list(APPEND HIP_HIPCC_FLAGS -std=c++17)
# Pass device library path for theRock nightly builds
if(DEFINED ENV{HIP_DEVICE_LIB_PATH})
file(TO_CMAKE_PATH "$ENV{HIP_DEVICE_LIB_PATH}" _hip_device_lib_path)
--- a/cmake/public/utils.cmake
+++ b/cmake/public/utils.cmake
@@ -254,6 +254,7 @@ function(caffe2_hip_binary_target target_name_or_src)
caffe2_binary_target(${target_name_or_src})
target_compile_options(${__target} PRIVATE ${HIP_CXX_FLAGS})
+ set_target_properties(${__target} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
target_include_directories(${__target} PRIVATE ${Caffe2_HIP_INCLUDE})
endfunction()

27
sci-ml/caffe2/files/caffe2-2.7.1-aotriton-fixes.patch
View File

@@ -1,27 +0,0 @@
Fix installation with aotriton
Upstream bug: https://github.com/pytorch/pytorch/issues/158109
--- a/cmake/External/aotriton.cmake
+++ b/cmake/External/aotriton.cmake
@@ -43,10 +43,6 @@ if(NOT __AOTRITON_INCLUDED)
# Note it is INSTALL"ED"
if(DEFINED ENV{AOTRITON_INSTALLED_PREFIX})
- install(DIRECTORY
- $ENV{AOTRITON_INSTALLED_PREFIX}/${CMAKE_INSTALL_LIBDIR}
- $ENV{AOTRITON_INSTALLED_PREFIX}/include
- DESTINATION ${__AOTRITON_INSTALL_DIR})
set(__AOTRITON_INSTALL_DIR "$ENV{AOTRITON_INSTALLED_PREFIX}")
message(STATUS "Using Preinstalled AOTriton at ${__AOTRITON_INSTALL_DIR}")
elseif(DEFINED ENV{AOTRITON_INSTALL_FROM_SOURCE})
--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
@@ -921,7 +921,7 @@ if(USE_ROCM)
set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
list(APPEND Caffe2_HIP_SRCS ${GENERATED_CXX_TORCH_CUDA})
hip_add_library(torch_hip ${Caffe2_HIP_SRCS})
- if(USE_FLASH_ATTENTION)
+ if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
target_link_libraries(torch_hip PRIVATE __caffe2_aotriton)
endif()
set(CUDA_LINK_LIBRARIES_KEYWORD)

53
sci-ml/caffe2/files/caffe2-2.9.0-cmake.patch
View File

@@ -1,53 +0,0 @@
--- a/.ci/pytorch/test_example_code/CMakeLists.txt 2025-10-24 18:25:10.584717735 +0200
+++ b/.ci/pytorch/test_example_code/CMakeLists.txt 2025-10-24 18:25:17.195639297 +0200
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
project(simple-torch-test)
find_package(Torch REQUIRED)
--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt 2025-10-24 18:26:46.165583599 +0200
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt 2025-10-24 18:27:00.085418425 +0200
@@ -4,7 +4,7 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
-cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
include(GNUInstallDirs)
--- a/aten/src/ATen/test/test_install/CMakeLists.txt 2025-10-24 18:27:36.180990366 +0200
+++ b/aten/src/ATen/test/test_install/CMakeLists.txt 2025-10-24 18:27:46.812864333 +0200
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
find_package(ATen REQUIRED)
include_directories(${ATEN_INCLUDE_DIR})
--- a/aten/src/ATen/nnapi/CMakeLists.txt 2025-10-24 18:28:58.948009238 +0200
+++ b/aten/src/ATen/nnapi/CMakeLists.txt 2025-10-24 18:29:11.040865888 +0200
@@ -1,6 +1,6 @@
# Define this to build the NNAPI binding out of tree.
if(PYTORCH_NNAPI_STANDALONE)
- cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
+ cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
project(pytorch_nnapi)
set(CMAKE_CXX_STANDARD 14 CACHE STRING "The C++ standard whose features are requested to build this target.")
--- a/android/pytorch_android_torchvision/CMakeLists.txt 2025-10-24 18:29:31.653621543 +0200
+++ b/android/pytorch_android_torchvision/CMakeLists.txt 2025-10-24 18:29:41.668502824 +0200
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
project(pytorch_vision_jni CXX)
set(CMAKE_CXX_STANDARD 17 CACHE STRING "The C++ standard whose features are requested to build this target.")
#_cmake_modify_IGNORE set(CMAKE_VERBOSE_MAKEFILE ON)
--- a/android/pytorch_android/CMakeLists.txt 2025-10-24 18:29:54.079355700 +0200
+++ b/android/pytorch_android/CMakeLists.txt 2025-10-24 18:30:39.793813797 +0200
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
option(BUILD_LITE_INTERPRETER "Master flag to build pytorch_jni_lite" ON)
message(
STATUS

28
sci-ml/caffe2/files/caffe2-2.9.0-rocm-distributed-link.patch
View File

@@ -1,28 +0,0 @@
USE_ROCM=ON USE_DISTRIBUTED=ON: fix *runtime* error: undefined symbol: rsmi_init
Upstream bug: https://github.com/pytorch/pytorch/issues/158725
--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
@@ -1058,6 +1058,12 @@ if(USE_ROCM)
)
endif()
+ if(USE_DISTRIBUTED)
+ list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS
+ rocm_smi64
+ )
+ endif()
+
# ---[ Kernel asserts
# Kernel asserts is disabled for ROCm by default.
# It can be turned on by turning on the env USE_ROCM_KERNEL_ASSERT to the build system.
--- a/cmake/public/LoadHIP.cmake
+++ b/cmake/public/LoadHIP.cmake
@@ -170,6 +170,7 @@ if(HIP_FOUND)
find_package_and_print_version(rocthrust REQUIRED)
find_package_and_print_version(hipsolver REQUIRED)
find_package_and_print_version(rocsolver REQUIRED)
+ find_package_and_print_version(rocm_smi REQUIRED)
# workaround cmake 4 build issue
if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.0.0")
message(WARNING "Work around hiprtc cmake failure for cmake >= 4")

1
sci-ml/pytorch/Manifest
View File

@@ -1,5 +1,4 @@
DIST composable_kernel-f1746955.tar.gz 7687217 BLAKE2B 973bf5b8eece2944e053c17eb3f8bf2d7f9a7744a651382c3d9677d6d4098f0107d64f3790899f74644681cdfb84e4989600c1e76bc1e96ee81579f870d55210 SHA512 dd03532d0487735d0ca6af5a37bf6a87917ab3ee81f70b14e29329d67a95f05bbccf18fb706ac9020918e7f6ac235e68b49164589fde073ff407a755daf26950
DIST flash-attention-2.7.4.gh.tar.gz 5841323 BLAKE2B 432999d763f2b3d732580ddfea5d3e01370351db0656546259a5e500a07516dd03c98828bfb55855dabe4adc651033b5d97ea4725ca46158b9970f0fbc662710 SHA512 05a4afb09e666f7404d6a3f8b5256e7bed6eba60a6f1bde2b7dbb96d318975f0b458c2521c7a38d88e97b6e4c27f29077cf787849daf82586e33f43a3d9a84b3
DIST pytorch-2.11.0.tar.gz 63504636 BLAKE2B d7227bd2d5475acf696bb48a3ca8cd8b7b6cd75fb99e64ebc71434e23817f94dea5164f48a176dcad1acd12186c33b02452582c1b41d6265b89cf54b29630670 SHA512 a5b27efdbce0506b49808e9dc5fbf1a56a22b1651b60236570823e1fe2c7395a0de9d3d7a01a568a201d593c52e7aa402e5d7645a54cb5932d5866e374ebf945
DIST pytorch-2.12.0.tar.gz 64740318 BLAKE2B e5e46a93422454da420563d0a8057c87d1ed4b27e0bd9b3b30e46b5e40cfbb489072144ac00267af2f085432d55e12adf750daf15ea165a1252449b00dca1ff9 SHA512 f5ab0f6933d88271f772b416f8c9b3b0d3e1ffaf8d00838b455206266b40d7c805e34003d84b01cddc7f1ad917dd72d6f79a05063b0962cbf223d9042cff3206
DIST pytorch-2.13.0.tar.gz 66369404 BLAKE2B 373df49b4f2d504aafedd948e34562696335d316ebcd497dbd86a54a3a3c39310d1595050cc2bcae858d0085e6a0c59ed5299962d6a2706441ab28c1699c4821 SHA512 0f7465835ae9206746d507e1b927afb94f8860248473f2d1c5d71ccd40defbd23d0e3e27a1b09deb907e753a7c9008352830170b53441701ade0916d68a42562

70
sci-ml/pytorch/pytorch-2.11.0.ebuild
View File

@@ -1,70 +0,0 @@
# Copyright 2022-2026 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2
EAPI=8
DISTUTILS_USE_PEP517=setuptools
PYTHON_COMPAT=( python3_{11..14} )
DISTUTILS_SINGLE_IMPL=1
DISTUTILS_EXT=1
inherit distutils-r1 prefix
DESCRIPTION="Tensors and Dynamic neural networks in Python"
HOMEPAGE="https://pytorch.org/"
SRC_URI="https://github.com/pytorch/${PN}/archive/refs/tags/v${PV}.tar.gz
-> ${P}.tar.gz"
LICENSE="BSD"
SLOT="0"
KEYWORDS="~amd64 ~arm64"
RESTRICT="test"
REQUIRED_USE=${PYTHON_REQUIRED_USE}
RDEPEND="
${PYTHON_DEPS}
~sci-ml/caffe2-${PV}[${PYTHON_SINGLE_USEDEP}]
$(python_gen_cond_dep '
dev-python/sympy[${PYTHON_USEDEP}]
dev-python/typing-extensions[${PYTHON_USEDEP}]
')
"
DEPEND="${RDEPEND}
$(python_gen_cond_dep '
dev-python/pyyaml[${PYTHON_USEDEP}]
')
"
PATCHES=(
"${FILESDIR}"/${PN}-2.9.0-dontbuildagain.patch
"${FILESDIR}"/${PN}-2.10.0-cpp-extension-multilib.patch
)
src_prepare() {
# Replace placeholders added by cpp-extension.patch
sed -e "s|%LIB_DIR%|$(get_libdir)|g" \
-i torch/utils/cpp_extension.py || die
# Set build dir for pytorch's setup
sed -e "/BUILD_DIR/s|build|/var/lib/caffe2/|" \
-i tools/setup_helpers/env.py || die
# Drop legacy from pyproject.toml
sed -e "/build-backend/s|:__legacy__||" \
-i pyproject.toml || die
distutils-r1_src_prepare
hprefixify tools/setup_helpers/env.py
}
python_compile() {
PYTORCH_BUILD_VERSION=${PV} \
PYTORCH_BUILD_NUMBER=0 \
USE_SYSTEM_LIBS=ON \
CMAKE_BUILD_DIR="${BUILD_DIR}" \
distutils-r1_python_compile develop sdist
}
python_install() {
USE_SYSTEM_LIBS=ON distutils-r1_python_install
}