Build with GCC, fixes build of 2.13.0 on aarch64

Improve paralellism on multisocket CPU
Current formula counts CPUs in first socket only. This change counts all physical cores in all sockets. I have system with two socket (each of them has 28 cores. As result the job is running with `-j 28` instead of `-j 112` and building at 25% utilization.
2026-08-21 15:32:41 -04:00 · 2026-08-18 22:42:50 +00:00 · 2026-08-18 13:51:28 -07:00 · 2026-08-15 17:49:49 -07:00
6 changed files with 147 additions and 40 deletions
--- a/.gitignore
+++ b/.gitignore
@ -41,3 +41,4 @@
 /kineto-b2103f7.tar.gz
 /pytorch-v2.12.0.tar.gz
 /v3.0.1.tar.gz
+/pytorch-v2.13.0.tar.gz
--- a/0001-pytorch-fmt.patch
+++ b/0001-pytorch-fmt.patch
@ -0,0 +1,54 @@
+From afdeb01359eecc67b573bcfb4f164ec81a1c7197 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 17 Aug 2026 15:59:39 -0700
+Subject: [PATCH] pytorch fmt
+
+---
+ cmake/Dependencies.cmake | 31 -------------------------------
+ 1 file changed, 31 deletions(-)
+
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index f7f938988c8e..d84eea5f1129 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1635,37 +1635,6 @@ endif()
+ # End ATen checks
+ #
+ 
+-# Install `fmtlib` header.
+-# This was the default behavior before version 12.0.0.
+-# Since PyTorch C API depends on it, make it available for projects that
+-# depend on PyTorch.
+-set(FMT_INSTALL ON)
+-set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+-set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
+-add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+-
+-# Disable compiler feature checks for `fmt`.
+-#
+-# CMake compiles a little program to check compiler features. Some of our build
+-# configurations (notably the mobile build analyzer) will populate
+-# CMAKE_CXX_FLAGS in ways that break feature checks. Since we already know
+-# `fmt` is compatible with a superset of the compilers that PyTorch is, it
+-# shouldn't be too bad to just disable the checks.
+-set_target_properties(fmt-header-only PROPERTIES INTERFACE_COMPILE_FEATURES "")
+-
+-# Keep fmt's header-only type layout stable across mixed C++ modes by forcing
+-# one no_unique_address spelling for all translation units.
+-if(MSVC AND NOT CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+-  set(_fmt_no_unique_address "[[msvc::no_unique_address]]")
+-else()
+-  set(_fmt_no_unique_address "[[no_unique_address]]")
+-endif()
+-target_compile_definitions(fmt PUBLIC "FMT_NO_UNIQUE_ADDRESS=${_fmt_no_unique_address}")
+-target_compile_definitions(fmt-header-only INTERFACE "FMT_NO_UNIQUE_ADDRESS=${_fmt_no_unique_address}")
+-unset(_fmt_no_unique_address)
+-
+-list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)
+-set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
+ 
+ # ---[ Kineto
+ # edge profiler depends on KinetoProfiler but it only does cpu
+-- 
+2.55.0
+
--- a/0001-pytorch-gloo.patch
+++ b/0001-pytorch-gloo.patch
@ -0,0 +1,27 @@
+From 1b1962613916961fade2ddb971c9ceeee74cc5c5 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 17 Aug 2026 08:16:49 -0700
+Subject: [PATCH] pytorch gloo
+
+---
+ cmake/Dependencies.cmake | 4 ++++
+ 1 file changed, 4 insertions(+)
+
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 81b673aabcd4..f7f938988c8e 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1277,6 +1277,10 @@ if(USE_GLOO)
+       if(NOT Gloo_FOUND)
+         message(FATAL_ERROR "Cannot find gloo")
+       endif()
+      set(Gloo_INCLUDE_DIRS /usr/include)
+      set(Gloo_NATIVE_LIBRARY /usr/lib64/libgloo.so)
+      set(Gloo_HIP_LIBRARY /usr/lib64/libgloo_hip.so)
+
+       message("Found gloo: ${Gloo_NATIVE_LIBRARY}, cuda lib: ${Gloo_CUDA_LIBRARY}, hip lib: ${Gloo_HIP_LIBRARY}")
+       message("Found gloo include directories: ${Gloo_INCLUDE_DIRS}")
+ 
+-- 
+2.55.0
+
--- a/0001-pytorch-xnnpack.patch
+++ b/0001-pytorch-xnnpack.patch
@ -0,0 +1,40 @@
+From ff660e83955efa54188d87c40804d991e31a8618 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 17 Aug 2026 07:34:20 -0700
+Subject: [PATCH] pytorch xnnpack
+
+---
+ cmake/Dependencies.cmake | 12 +++++++-----
+ 1 file changed, 7 insertions(+), 5 deletions(-)
+
+diff --git a/cmake/Dependencies.cmake b/cmake/Dependencies.cmake
+index 2baa187e69ac..81b673aabcd4 100644
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -611,16 +611,18 @@ if(USE_XNNPACK AND NOT USE_SYSTEM_XNNPACK)
+   list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK microkernels-prod)
+ elseif(NOT TARGET XNNPACK AND USE_SYSTEM_XNNPACK)
+   add_library(XNNPACK SHARED IMPORTED)
+-  add_library(microkernels-prod SHARED IMPORTED)
+#  add_library(microkernels-prod SHARED IMPORTED)
+   find_library(XNNPACK_LIBRARY XNNPACK)
+-  find_library(microkernels-prod_LIBRARY microkernels-prod)
+#  find_library(microkernels-prod_LIBRARY microkernels-prod)
+   set_property(TARGET XNNPACK PROPERTY IMPORTED_LOCATION "${XNNPACK_LIBRARY}")
+-  set_property(TARGET microkernels-prod PROPERTY IMPORTED_LOCATION "${microkernels-prod_LIBRARY}")
+-  if(NOT XNNPACK_LIBRARY OR NOT microkernels-prod_LIBRARY)
+#  set_property(TARGET microkernels-prod PROPERTY IMPORTED_LOCATION "${microkernels-prod_LIBRARY}")
+#  if(NOT XNNPACK_LIBRARY OR NOT microkernels-prod_LIBRARY)
+  if(NOT XNNPACK_LIBRARY)
+     message(FATAL_ERROR "Cannot find XNNPACK")
+   endif()
+   message("-- Found XNNPACK: ${XNNPACK_LIBRARY}")
+-  list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK microkernels-prod)
+#  list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK microkernels-prod)
+    list(APPEND Caffe2_DEPENDENCY_LIBS XNNPACK)
+ endif()
+ 
+ # ---[ Vulkan deps
+-- 
+2.55.0
+
--- a/python-torch.spec
+++ b/python-torch.spec
@ -15,7 +15,7 @@
 %global miniz_version 3.0.2
 %global pybind11_version 3.0.1
 %else
-%global pypi_version 2.12.0
+%global pypi_version 2.13.0
 %global flatbuffers_version 24.12.23
 %global miniz_version 3.0.2
 %global pybind11_version 3.0.1
@ -33,7 +33,7 @@
 %endif

 # For testing distributed+rccl etc.
-%bcond_with gloo
+%bcond_without gloo
 %bcond_without mpi
 %bcond_without tensorpipe

@ -51,8 +51,8 @@
 # In fedora, not in rhel/epel or requires a newer version
 %if 0%{?fedora}
 %bcond_without eigen3
-%bcond_with onnx
-%bcond_with protobuf
+%bcond_without onnx
+%bcond_without protobuf
 %bcond_with setuptools
 %bcond_without sympy
 %else
@ -87,11 +87,15 @@ Patch2: 0001-Fix-struct.pack-polyfill-signature-mismatch-on-Pytho.patch
 %else
 Source0:        %{forgeurl}/releases/download/v%{version}/pytorch-v%{version}.tar.gz

+Patch1:         0001-pytorch-xnnpack.patch
+Patch2:         0001-pytorch-gloo.patch
+Patch3:         0001-pytorch-fmt.patch
+
 Source1001:     inject.py

 # Problems with python 3.15
-Patch1: 0001-Fix-functools.reduce-polyfill-signature-mismatch-on-.patch
-Patch2: 0001-Fix-struct.pack-polyfill-signature-mismatch-on-Pytho.patch
+# Patch1: 0001-Fix-functools.reduce-polyfill-signature-mismatch-on-.patch
+# Patch2: 0001-Fix-struct.pack-polyfill-signature-mismatch-on-Pytho.patch

 %endif
 Source1:        https://github.com/google/flatbuffers/archive/refs/tags/v%{flatbuffers_version}.tar.gz
@ -155,13 +159,7 @@ Source120:       https://github.com/meta-pytorch/MSLK/archive/%{mslk_commit}/MSL

 %global pt_arches x86_64 aarch64
 ExclusiveArch:  %pt_arches
-# Switch to clang temporarily on aarch64, GCC 16 + PyTorch 2.9.1 has problems
-# See rhbz#2431943 for more info
-%ifarch aarch64
-%global toolchain clang
-%else
 %global toolchain gcc
-%endif
 %global _lto_cflags %nil

 BuildRequires:  cmake
@ -196,7 +194,7 @@ BuildRequires:  onnx-devel
 BuildRequires:  openmpi-devel
 %endif
 %if %{with protobuf}
-BuildRequires:  protobuf-devel < 4
+BuildRequires:  protobuf-devel
 %endif
 BuildRequires:  sleef-devel
 BuildRequires:  valgrind-devel
@ -208,6 +206,8 @@ BuildRequires:  fxdiv-devel
 BuildRequires:  psimd-devel
 BuildRequires:  xnnpack-devel = 0.0^git20240814.312eb7e

+BuildRequires:  zlib-ng-compat-static
+
 BuildRequires:  python3-devel
 BuildRequires:  python3dist(filelock)
 BuildRequires:  python3dist(jinja2)
@ -258,6 +258,7 @@ BuildRequires:  rocm-runtime-devel
 BuildRequires:  rocm-rpm-macros
 BuildRequires:  rocsolver-devel
 BuildRequires:  rocm-smi-devel
+BuildRequires:  rocshmem-devel
 BuildRequires:  rocthrust-devel
 BuildRequires:  roctracer-devel

@ -497,7 +498,7 @@ sed -i -e 's@set(USE_MIMALLOC ON)@set(USE_MIMALLOC OFF)@' CMakeLists.txt
 mv third_party/miniz-%{miniz_version} .
 #
 # setup.py depends on this script
-mv third_party/build_bundled.py .
+# mv third_party/build_bundled.py .

 # Need the just untarred flatbuffers/flatbuffers.h
 mv third_party/flatbuffers .
@ -535,7 +536,7 @@ mv third_party/googletest .
 # Remove everything
 rm -rf third_party/*
 # Put stuff back
-mv build_bundled.py third_party
+# mv build_bundled.py third_party
 mv miniz-%{miniz_version} third_party
 mv flatbuffers third_party
 mv pybind11 third_party
@ -625,7 +626,7 @@ sed -i -e 's@USE_MSLK_DEFAULT ON@USE_MSLK_DEFAULT OFF@' CMakeLists.txt
 #
 %ifarch x86_64
 # Real cores, No hyperthreading
-COMPILE_JOBS=`cat /proc/cpuinfo | grep -m 1 'cpu cores' | awk '{ print $4 }'`
+COMPILE_JOBS=`lscpu -p=CORE | grep -v '^#' | sort -u |wc -l`
 %else
 # cpuinfo format varies on other arches, fall back to nproc
 COMPILE_JOBS=`nproc`
@ -698,7 +699,7 @@ export USE_SYSTEM_EIGEN_INSTALL=ON
 export USE_SYSTEM_ONNX=ON
 %endif
 export USE_SYSTEM_PYBIND11=OFF
-export USE_SYSTEM_LIBS=OFF
+export USE_SYSTEM_LIBS=ON
 export USE_SYSTEM_NCCL=OFF
 export USE_TENSORPIPE=OFF
 export USE_XNNPACK=OFF
@ -708,7 +709,7 @@ export USE_SYSTEM_CPUINFO=ON
 export USE_SYSTEM_FP16=ON
 export USE_SYSTEM_FXDIV=ON
 export USE_SYSTEM_PSIMD=ON
-export USE_SYSTEM_XNNPACK=OFF
+export USE_SYSTEM_XNNPACK=ON

 export USE_DISTRIBUTED=ON
 %if %{with tensorpipe}
@ -717,7 +718,7 @@ export TP_BUILD_LIBUV=OFF
 %endif

 %if %{with gloo}
-export USE_GLOO=ON
+export USE_GLOO=OFF
 export USE_SYSTEM_GLOO=ON
 %endif
 %if %{with mpi}
--- a/26
+++ b/26
@ -1,29 +1,13 @@
-SHA512 (pytorch-v2.7.0.tar.gz) = 17e875a66f1669901f5f770c9d829ba5bfa3967296cfb71550e8a92507181db742548eaf7cc9a2c478c4b91e366f27cc480e2e1bbb328db8501d30e1649839e6
-SHA512 (v23.3.3.tar.gz) = 4066c94f2473c7ea16917d29a613e16f840a329089c88e0bdbdb999aef3442ba00abfd2aa92266fa9c067e399dc88e6f0ccac40dc151378857e665638e78bbf0
-SHA512 (v2.13.6.tar.gz) = 497c25b33b09a9c42f67131ab82e35d689e8ce089dd7639be997305ff9a6d502447b79c824508c455d559e61f0186335b54dd2771d903a7c1621833930622d1a
+SHA512 (pytorch-v2.13.0.tar.gz) = 8d192f2d4ef92e4f83cef86d97f4f23fceec4fd4eaa2ff1d99267250d64785317e918617ef9b80c7cb1dfa209614444b4f427caa7a48d10911cbd2caeac1655f
+SHA512 (v24.12.23.tar.gz) = f97762ba41b9cfef648e93932fd789324c6bb6ebc5b7aeca8185c9ef602294b67d73aea7ae371035579a1419cbfbeba7c3e88b31b5a5848db98f5e8a03b982b1
+SHA512 (v3.0.1.tar.gz) = c17e6d6a78c38e760864b390ac2aa7df6a94ca53acb2e8be71f0d63d611b738fa20a16946c98a93fbfcad56cb0346ebf247bbe41c6f5171c6ce68397b1e5c4db
 SHA512 (tensorpipe-52791a2.tar.gz) = 1e5faf17a7236c5506c08cb28be16069b11bb929bbca64ed9745ce4277d46739186ab7d6597da7437d90ed2d166d4c37ef2f3bceabe8083ef3adbb0e8e5f227e
 SHA512 (v1.41.0.tar.gz) = bb08a1970a10e8d9571ffea3d021643de30ec212cd51317b98d6cf0cfe55d6877992921fb01d1188a6d466687335b77885685d924f8cb7200a0bec30eee05c65
 SHA512 (libnop-910b558.tar.gz) = 74c5324eaa1b6b2ac8dfef94c835b5c5b044625f8e5efe3522470b1ecc4798ff43d344a013cee2f6901e83267c6167072947b754e63f1552ae7044cffe234c36
 SHA512 (v1.14.2.tar.gz) = 97635bbaf6dd567c201451dfaf7815b2052fe50d9bccc97aade86cfa4a92651374d167296a5453031b2681dc302806a289bca011a9e79ddc381a17d6118971d7
-SHA512 (cpp-httplib-3b6597b.tar.gz) = 8f1090658c498d04f14fec5c2f301847b1f3360bf92b18d82927643ee04ab61a6b274733a01c7850f9c030205120d674d1d961358d49fdd15636736fb8704f55
-SHA512 (kineto-be13176.tar.gz) = 41a08c7da9eea7d12402f80a5550c9d4df79798719cc52b12a507828c8c896ba28a37c35d8adf809ca72589e1d84965d5ef6dd01f3f8dc1c803c5ed67b03a43a
-SHA512 (pytorch-a1cb3cc.tar.gz) = 92bf8b2c2ef0b459406b60169ecebdc50652c75943e3d6087e4d261f6e308dbad365529561e0f07ea3f0b71790efb68b5e4ab2f44e270462097208d924dc2d95
-SHA512 (v24.12.23.tar.gz) = f97762ba41b9cfef648e93932fd789324c6bb6ebc5b7aeca8185c9ef602294b67d73aea7ae371035579a1419cbfbeba7c3e88b31b5a5848db98f5e8a03b982b1
-SHA512 (kineto-5e75018.tar.gz) = 921b96a56e01d69895b79e67582d8977ed6f873573ab41557c5d026ada5d1f6365e4ed0a0c6804057c52e92510749fc58619f554a164c1ba9d8cd13e789bebd0
-SHA512 (pytorch-v2.8.0.tar.gz) = 791e658eab87fb957f025558cb9f925078d2426ab7b6f60771d9841dfb691f67d905ba1330a800008efe7c938b6c69bdc52232bccfe8d4860e795a532cd69d28
+SHA512 (cpp-httplib-bd95e67.tar.gz) = 17c7b1d1a3750b85ebfefcc31a2fe2cb538373b140a31058d5003a018601bbde96e82bd73b99708856c0ebecd910ac4aee5420b829cb444e3aac3ee5764dcff7
+SHA512 (kineto-b2103f7.tar.gz) = 27d9f6a8b27434e83d26ed182df1fecec1d02fc26906995e44155562a730f998f527aad3d6c8a37212255c5e2f86a607a9648edf62861f56eb2ea512c4452908
 SHA512 (v1.18.0.tar.gz) = 2f38664947c8d1efc40620a7c1b1953d2aa4b0a37b67c4886b86e77c1d697363c26413413ddda8eabc545892fb1bcb43afc7e93e62f0901527524a2727e1ea8d
-SHA512 (pytorch-715dca6.tar.gz) = 09c9aae54fab3eb17901fc3226fece1c13f41cb8e45a2cb066021823abeb8d27c340993088e01d8e55bb37ed5f94334ec31e6c539cddfacbad157abd27c5e907
-SHA512 (pytorch-fd36458.tar.gz) = acbb7475b92ad4a8e8d779f3745da22d8438e4c5ef2d6e76d71c987789f2752c8aef7022c87c9a74640fe4f9c1f1a61a3f12a796f63b1e6be24da8e5aacf37dc
-SHA512 (pytorch-0fabc3b.tar.gz) = 2e87975de0bf6f3dcede168b379e1928712bca16170c2a8ee7d63459f53086c01baac05e0763e4d5d28cdaf1c7d8912225ee06adeff96ead4f6f456ee174b341
-SHA512 (pytorch-v2.9.0.tar.gz) = ae989e3a7fe30f9ea90944dc25e21ca92f2a94ee40d8de974a168c292d82c16ee8920624eff91a85755469ad05473dce0f85893e3ed7794ec5c6bdd89cbd2023
-SHA512 (pytorch-v2.9.1.tar.gz) = 88de0289fa2760abd69bef505b5ae3b6d7ff176b415cbb31bbc89ce5476a3800b322a97c4490f270f8b89657aff931bf9a5516202b268e0bb8b1f63dbb87b34a
 SHA512 (v3.19.6.tar.gz) = 8f92242f2be8e1bbfba41341c87709ad91ad83b8b3e3df88bb430411541d3399295f49291fd52b50e3487b0fce33181cb4d175685fd25aac72adfaee26a612d4
 SHA512 (v80.9.0.tar.gz) = e67c6c5e74691e65ecbf24fd19c1da545434ac1674f37c917f5ecc282a00bffc62d64164a885d19f177e99a3fa53db196e2be1cf0aa6811027e61d05119682da
-SHA512 (pytorch-v2.10.0.tar.gz) = 64f5c62ac9ec7189ab27a787271e2d993d3230638b43792bbf6576e8ada840323d86ebfc4262d951266281623518cca5ba87f77ca61e165fef6a670fdf5b3c50
-SHA512 (pytorch-v2.11.0.tar.gz) = e67d68245a19e65737c19f97b6d4ddd72051303abb17a734e09f1ba7ecfb3bf6a5546c826820890ad0ddfb2916b1f798175f341bda0cc3c8a83b3b852b077f3b
 SHA512 (MSLK-3d332d1.tar.gz) = f6bd99ec7a79321692c088e622dd2ea8f1355b137014b33e75cb7ff83922fdd07ebacf5920e2e343a48fccd74315c11861ac5c2fff29e2f2481875ecb1ae57cf
-SHA512 (cpp-httplib-bd95e67.tar.gz) = 17c7b1d1a3750b85ebfefcc31a2fe2cb538373b140a31058d5003a018601bbde96e82bd73b99708856c0ebecd910ac4aee5420b829cb444e3aac3ee5764dcff7
-SHA512 (kineto-7a731b6.tar.gz) = f15347cdb210b887c1c9c9e9d91556f48b5b3fc662dce0bfa6086bce38673c568f0fc385cdd1b16a742cc709d4db5798a18c6400f24bd57f6a1b4a2eaf303d47
-SHA512 (kineto-b2103f7.tar.gz) = 27d9f6a8b27434e83d26ed182df1fecec1d02fc26906995e44155562a730f998f527aad3d6c8a37212255c5e2f86a607a9648edf62861f56eb2ea512c4452908
-SHA512 (pytorch-v2.12.0.tar.gz) = 00f900945c3a7dc33fe61ece65341d7995c13f6134b714bc246f43b06f7d091b5952e8cb3febda6931cb769c96189e6b646c179ea59601ead834fb007f0788f2
-SHA512 (v3.0.1.tar.gz) = c17e6d6a78c38e760864b390ac2aa7df6a94ca53acb2e8be71f0d63d611b738fa20a16946c98a93fbfcad56cb0346ebf247bbe41c6f5171c6ce68397b1e5c4db
Author	SHA1	Message	Date
Alexander F. Lent	74ef34969c	Build with GCC, fixes build of 2.13.0 on aarch64	2026-08-21 15:32:41 -04:00
Miroslav Suchý	749d91ee2c	Improve paralellism on multisocket CPU Current formula counts CPUs in first socket only. This change counts all physical cores in all sockets. I have system with two socket (each of them has 28 cores. As result the job is running with `-j 28` instead of `-j 112` and building at 25% utilization.	2026-08-18 22:42:50 +00:00
Tom Rix	087f2caeb5	Update to 2.13 Signed-off-by: Tom Rix <Tom.Rix@amd.com>	2026-08-18 13:51:28 -07:00
Tom Rix	a3aaf8df1b	Rebuild for ROCm 7.14 Signed-off-by: Tom Rix <Tom.Rix@amd.com>	2026-08-15 17:49:49 -07:00