python-triton: keep the nvidia C++ backend in the cmake build

Dropping nvidia from TRITON_CODEGEN_BACKENDS does not build the NVGPU/NVWS dialect TableGen output under third_party/nvidia, but Triton core hard- depends on it (TritonGPUTransforms via Passes.h, TritonInstrumentToLLVM), so the wheel build failed with fatal error: 'nvidia/include/Dialect/NVGPU/IR/Dialect.h.inc' file not found fatal error: 'nvidia/include/Dialect/NVWS/IR/Dialect.h.inc' file not found Keep the nvidia C++ libraries in the cmake build (they only need the in-tree LLVM NVPTX target, no CUDA SDK, and stay offline) while still packaging only the AMD Python backend, so neither ptxas nor the proprietary libdevice.10.bc lands in the RPM. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
python-triton: add pybind11 as deps
2026-06-11 10:00:25 +08:00 · 2026-06-11 03:00:51 +08:00 · 2026-06-11 02:56:10 +08:00 · 2026-06-10 13:29:01 +08:00 · 2026-06-08 15:58:11 +08:00 · 2026-06-08 15:40:51 +08:00
78 changed files with 5032 additions and 1788 deletions
@@ -0,0 +1,38 @@
+From b8b4a6bcfe35ba9539a120cfd16573123ddd9241 Mon Sep 17 00:00:00 2001
+From: "Sv. Lockal" <lockalsash@gmail.com>
+Date: Mon, 15 Dec 2025 03:46:35 +0800
+Subject: [PATCH] Fix compilation with libdrm-2.4.130
+
+Fix error: redefinition of 'struct drm_color_ctm_3x4'.
+
+drm_color_ctm_3x4 structure is now defined in https://github.com/torvalds/linux/commit/e5719e7f19009d4fbedf685fc22eec9cd8de154f#diff-4c51fb416ec7cc69566cd7b795ee57eb070aa1006ad65d6962081f039ffb2718
+
+As this structure is unused and not a part of amdsmi public interface,
+it is safe to remove it.
+---
+ include/amd_smi/impl/amdgpu_drm.h | 9 ---------
+ 1 file changed, 9 deletions(-)
+
+diff --git a/include/amd_smi/impl/amdgpu_drm.h b/include/amd_smi/impl/amdgpu_drm.h
+index b56a5ac4b20a..0e483d13b382 100644
+--- a/include/amd_smi/impl/amdgpu_drm.h
+++ b/include/amd_smi/impl/amdgpu_drm.h
+@@ -1625,15 +1625,6 @@ struct drm_amdgpu_info_uq_metadata {
+ #define AMDGPU_FAMILY_GC_11_5_0			150 /* GC 11.5.0 */
+ #define AMDGPU_FAMILY_GC_12_0_0			152 /* GC 12.0.0 */
+ 
+-/* FIXME wrong namespace! */
+-struct drm_color_ctm_3x4 {
+-	/*
+-	 * Conversion matrix with 3x4 dimensions in S31.32 sign-magnitude
+-	 * (not two's complement!) format.
+-	 */
+-	__u64 matrix[12];
+-};
+-
+ #if defined(__cplusplus)
+ }
+ #endif
+-- 
+2.51.1
+
@@ -0,0 +1,36 @@
+From 21afd2c2d58b8c7895df47f6a16a0781a7f0024a Mon Sep 17 00:00:00 2001
+From: Sakura286 <sakura286@outlook.com>
+Date: Sat, 6 Jun 2026 20:38:13 +0800
+Subject: [PATCH 1/2] Disable goamdsmi_shim when ESMI is off
+
+The Go shim in goamdsmi_shim/smiwrapper/amdsmi_go_shim.c calls CPU-only
+APIs such as amdsmi_get_cpu_core_energy, amdsmi_get_threads_per_core
+and amdsmi_get_processor_handles_by_type. In include/amd_smi/amdsmi.h
+these are all guarded by `#ifdef ENABLE_ESMI_LIB`, so when ESMI is
+disabled (e.g. on riscv64 / non-x86 architectures) the declarations
+disappear and the shim fails to build with implicit-function-declaration
+errors.
+---
+ CMakeLists.txt | 6 ++++--
+ 1 file changed, 4 insertions(+), 2 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 1b7375f..013c1ad 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -363,8 +363,10 @@ install(
+     PATTERN "build*" EXCLUDE
+     PATTERN ".cache*" EXCLUDE)
+ 
+-# Make for goamdsmi_shim library
+-add_subdirectory(goamdsmi_shim)
+# The Go shim uses CPU APIs gated by ENABLE_ESMI_LIB; only build it when ESMI is on
+if(ENABLE_ESMI_LIB)
+    add_subdirectory(goamdsmi_shim)
+endif()
+ 
+ #Debian package specific variables
+ set(CPACK_DEBIAN_PACKAGE_RECOMMENDS "python3-argcomplete, libdrm-dev, libdrm-amdgpu-dev")
+-- 
+2.53.0
+
@@ -0,0 +1,77 @@
+From 41740f15ede6e04e46ff736bcc85ca8fd1aae641 Mon Sep 17 00:00:00 2001
+From: Sakura286 <sakura286@outlook.com>
+Date: Sat, 6 Jun 2026 21:55:37 +0800
+Subject: [PATCH 2/2] Tolerate missing CPU/E-SMI symbols on non-x86_64
+
+With ENABLE_ESMI_LIB=OFF (non-x86_64), libamd_smi.so omits the CPU API,
+but the ctypesgen wrapper binds every symbol at import time, so the
+missing CPU symbols make `import amdsmi` fail with AttributeError.
+
+Wrap the loaded CDLL in a proxy so missing symbols resolve to a lazy
+stub that only raises when actually called. `import amdsmi` then works
+and GPU consumers (e.g. PyTorch with ROCM EP, which never calls the CPU
+API) are unaffected. The library object is wrapped instead of ctypes.CDLL
+itself so callers hooking ctypes.CDLL keep working.
+---
+ py-interface/amdsmi_wrapper.py | 38 +++++++++++++++++++++++++++++++++-
+ 1 file changed, 37 insertions(+), 1 deletion(-)
+
+diff --git a/py-interface/amdsmi_wrapper.py b/py-interface/amdsmi_wrapper.py
+index 99ed017..0626b72 100644
+--- a/py-interface/amdsmi_wrapper.py
+++ b/py-interface/amdsmi_wrapper.py
+@@ -176,6 +176,42 @@ from pathlib import Path
+ # 3. Relative to amdsmi_wrapper.py
+ #    - parent directory
+ #    - current directory
+class _AmdsmiMissingSymbol:
+    # Placeholder for a symbol absent from libamd_smi.so (e.g. the CPU/E-SMI
+    # API when the library is built with ENABLE_ESMI_LIB=OFF on non-x86_64).
+    # It accepts the restype/argtypes the wrapper assigns at import time and
+    # only raises if the symbol is ever actually called.
+    def __init__(self, name):
+        object.__setattr__(self, "_amdsmi_name", name)
+
+    def __setattr__(self, key, value):
+        object.__setattr__(self, key, value)
+
+    def __call__(self, *args, **kwargs):
+        name = object.__getattribute__(self, "_amdsmi_name")
+        raise NotImplementedError(
+            "amdsmi symbol " + repr(name) + " is unavailable: it is not "
+            "exported by this build of libamd_smi.so")
+
+
+class _AmdsmiTolerantLib:
+    # Proxy around the loaded CDLL so that binding a symbol the library does
+    # not export resolves to a stub instead of raising AttributeError at
+    # import time. We wrap the already-constructed library object rather than
+    # the ctypes.CDLL() call so that callers hooking ctypes.CDLL (e.g.
+    # PyTorch's libamd_smi.so loader) keep working.
+    def __init__(self, lib):
+        object.__setattr__(self, "_amdsmi_lib", lib)
+
+    def __getattr__(self, name):
+        try:
+            return getattr(object.__getattribute__(self, "_amdsmi_lib"), name)
+        except AttributeError:
+            if name.startswith("__"):
+                raise
+            return _AmdsmiMissingSymbol(name)
+
+
+ def find_smi_library():
+     err = OSError("Could not load libamd_smi.so")
+     possible_locations = []
+@@ -194,7 +230,7 @@ def find_smi_library():
+     for location in possible_locations:
+         try:
+             lib = ctypes.CDLL(location)
+-            return lib, location
+            return _AmdsmiTolerantLib(lib), location
+         except OSError as e:
+             err = e
+             continue
+-- 
+2.53.0
+
@@ -0,0 +1,175 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%bcond test 0
+%if %{with test}
+%global build_test ON
+%else
+%global build_test OFF
+%endif
+
+%global rocm_release 7.2
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# esmi_ib_library is not suitable for packaging
+# https://github.com/amd/esmi_ib_library/issues/13
+# This tag was chosen by the amdsmi project because 4.0+ introduced variables
+# not found in the upstream kernel.
+%global esmi_ver 4.2
+%global pkg_library_version 26
+
+Name:           amdsmi
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        AMD System Management Interface
+License:        MIT AND (GPL-2.0-only WITH Linux-syscall-note) AND NSCA
+# Main license is MIT
+#
+# This file is GPL-2.0:
+# include/amd_smi/impl/amd_hsmp.h
+# esmi_ib_library/include/asm/amd_hsmp.h
+# Both carry: SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note
+#
+# NSCA covers the bundled esmi_ib_library
+Url:            https://github.com/ROCm/rocm-systems
+#!RemoteAsset:  sha256:23c31cd787d86ee35c82746fcde705eacc46517815110376f28417909ef46406
+Source0:        %{url}/releases/download/rocm-%{version}/%{name}.tar.gz
+#!RemoteAsset:  sha256:de19d222d09e2171f47f8bbd6608e5648bd547c82543379bb8fb5ed2e379e141
+Source1:        https://github.com/amd/esmi_ib_library/archive/refs/tags/esmi_pkg_ver-%{esmi_ver}.tar.gz
+BuildSystem:    cmake
+
+# Support libdrm 2.4.130+
+# https://github.com/ROCm/amdsmi/pull/165
+Patch0:         0001-Fix-compilation-with-libdrm-2.4.130.patch
+# -DENABLE_ESMI_LIB=OFF is not enough.
+# Goamdshim references CPU/ESMI-only APIs; only build it when ESMI is on
+Patch1:         2001-Disable-goamdsmi_shim-when-ESMI-is-off.patch
+# Without ESMI (non-x86_64) libamd_smi.so omits the CPU API; let the ctypesgen
+# wrapper tolerate the missing symbols so `import amdsmi` still works
+Patch2:         2002-Tolerate-missing-CPU-E-SMI-symbols-on-non-x86_64.patch
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DBUILD_TESTS=%{build_test}
+BuildOption(conf):  -DCMAKE_SKIP_INSTALL_RPATH=TRUE
+%ifnarch x86_64
+BuildOption(conf):  -DENABLE_ESMI_LIB=OFF
+%endif
+
+BuildRequires:  cmake
+%if %{with test}
+BuildRequires:  cmake(GTest)
+%endif
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(libdrm)
+BuildRequires:  pkgconfig(libdrm_amdgpu)
+BuildRequires:  pkgconfig(python3)
+
+Requires:       python3dist(pyyaml)
+
+%description
+The AMD System Management Interface Library, or AMD SMI library, is a C
+library for Linux that provides a user space interface for applications
+to monitor and control AMD devices.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep
+%autosetup -p1 -N -n %{name}
+%patch 0 -p1
+%patch 1 -p1
+%ifnarch x86_64
+%patch 2 -p1
+%endif
+
+# ESMI - EPYC System Management Interface
+# esmi_ib_library uses x86-only cpuid.h; guard it for non-x86 builds
+%ifarch x86_64
+tar xf %{SOURCE1}
+mv esmi_ib_library-* esmi_ib_library
+mv esmi_ib_library/License.txt esmi_ib_library_License.txt
+# The esmi version check uses git tags, but we use tar's without git files.
+# Just inject in the tag that we've pulled into the version check:
+sed -i 's/NOT latest_esmi_tag/NOT "esmi_pkg_ver-%{esmi_ver}"/' CMakeLists.txt
+%endif
+
+# /usr/libexec/amdsmi_cli/BDF.py:126: SyntaxWarning: invalid escape sequence '\.'
+sed -i -e 's@bdf_regex = "@bdf_regex = r"@' amdsmi_cli/BDF.py
+
+# Fix script shebang
+sed -i -e 's@env python3@python3@' amdsmi_cli/*.py
+
+%install -a
+mkdir -p %{buildroot}%{python3_sitearch}
+mv %{buildroot}%{_datadir}/amdsmi %{buildroot}%{python3_sitearch}
+mv %{buildroot}%{_datadir}/pyproject.toml %{buildroot}%{python3_sitearch}/amdsmi/
+
+# W: unstripped-binary-or-object .../amdsmi/libamd_smi.so
+# Does an explicit open, so can not just rm it; strip it instead
+strip %{buildroot}%{python3_sitearch}/amdsmi/*.so
+# E: non-executable-script .../amdsmi_cli/amdsmi_cli_exceptions.py 644 /usr/bin/env python3
+chmod a+x %{buildroot}%{_libexecdir}/amdsmi_cli/amdsmi_*.py
+
+rm -rf %{buildroot}%{_datadir}/example
+rm -rf %{buildroot}%{_datadir}/amd_smi/example
+rm -f %{buildroot}%{_datadir}/_version.py
+rm -f %{buildroot}%{_datadir}/amd_smi/_version.py
+rm -f %{buildroot}%{_datadir}/setup.py
+rm -f %{buildroot}%{_datadir}/amd_smi/setup.py
+rm -f %{buildroot}%{_docdir}/amd_smi-asan/LICENSE.txt
+rm -f %{buildroot}%{_docdir}/amd-smi-lib/LICENSE.txt
+rm -f %{buildroot}%{_docdir}/amd-smi-lib/README.md
+rm -rf %{buildroot}%{_docdir}/amd-smi-lib/copyright
+
+if [ -e %{buildroot}%{_datadir}/amd_smi/tests ]; then
+    mkdir -p %{buildroot}%{_datadir}/amdsmi
+    mv %{buildroot}%{_datadir}/amd_smi/tests %{buildroot}%{_datadir}/amdsmi/
+fi
+
+%files
+%doc README.md
+%license LICENSE
+%{_bindir}/amd-smi
+%{_libdir}/libamd_smi.so.%{pkg_library_version}{,.*}
+%{_libexecdir}/amdsmi_cli
+%{python3_sitearch}/amdsmi
+
+%ifarch x86_64
+%license esmi_ib_library_License.txt
+%{_libdir}/libgoamdsmi_shim64.so.1{,.*}
+%endif
+
+%files devel
+%{_includedir}/amd_smi/
+%{_libdir}/cmake/amd_smi/
+%{_libdir}/libamd_smi.so
+
+%ifarch x86_64
+%{_includedir}/*.h
+%{_libdir}/libgoamdsmi_shim64.so
+%endif
+
+%if %{with test}
+%files test
+%{_datadir}/amdsmi/
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,36 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+Name:           fplus
+Version:        0.2.28
+Release:        %autorelease
+Summary:        Helps you write concise and readable C++ code
+Url:            https://github.com/Dobiasd/FunctionalPlus
+License:        BSL-1.0
+#!RemoteAsset:  sha256:8864a3e9bebde6ebed71b49ac2a036cedf9ae0f02ce758bc28c21e6a2ae15803
+Source0:         %{url}/archive/v%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildRequires:  cmake
+
+%description
+FunctionalPlus is a small header-only library supporting you in
+reducing code noise and in dealing with only one single level
+of abstraction at a time. By increasing brevity and maintainability
+of your code it can improve productivity (and fun!) in the long
+run. It pursues these goals by providing pure and easy-to-use
+functions that free you from implementing commonly used flows of
+control over and over again.
+
+%files
+%doc README.md
+%license LICENSE
+%{_includedir}/fplus/
+%{_libdir}/cmake/FunctionalPlus/
+
+%changelog
+%autochangelog
@@ -0,0 +1,40 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_release 7.1
+%global rocm_patch   1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+Name:           half
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm half-precision floating point library
+License:        MIT
+Url:            https://github.com/ROCm/half
+#!RemoteAsset:  sha256:1b5de9e50513560265a79022fd74322b77216f9bf938be688709a8e7d1d8d09d
+Source0:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildRequires:  cmake
+BuildRequires:  rocm-cmake
+
+%description
+half is a C++ header-only library providing an IEEE-754 conformant
+half-precision floating point type along with arithmetic operators,
+type conversions, and common mathematical functions. It is part of
+the ROCm software stack.
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/half/LICENSE.txt
+
+%files
+%license LICENSE.txt
+%doc README.txt
+%{_includedir}/half/
+
+%changelog
+%autochangelog
@@ -1,44 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global rocm_version 7.1.1
-
-Name:           hipblas-common
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        Common files shared by hipBLAS and hipBLASLt
-License:        MIT
-Url:            https://github.com/ROCm/hipBLAS-common
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildArch:      noarch
-BuildSystem:    cmake
-
-BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake
-
-%description
-%summary
-
-%package        devel
-Summary:        Libraries and headers for %{name}
-Provides:       %{name}-static = %{version}-%{release}
-
-%description    devel
-%{summary}
-
-%install -a
-rm -f %{buildroot}%{_prefix}/share/doc/hipblas-common/LICENSE.md
-
-%files devel
-%license LICENSE.md
-%{_includedir}/%{name}
-%{_libdir}/cmake/%{name}
-
-%changelog
-%{?autochangelog}
@@ -1,91 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global rocm_version 7.1.1
-
-%bcond test 0
-
-Name:           hipblas
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        ROCm BLAS marshalling library
-License:        MIT
-Url:            https://github.com/ROCm/hipBLAS
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-BuildRequires:  clang
-BuildRequires:  clang-tools-extra
-BuildRequires:  cmake
-BuildRequires:  cmake(amd_comgr)
-BuildRequires:  cmake(hip)
-BuildRequires:  cmake(hipblas-common)
-BuildRequires:  cmake(hsa-runtime64)
-BuildRequires:  cmake(rocblas)
-BuildRequires:  cmake(rocsolver)
-BuildRequires:  compiler-rt
-BuildRequires:  gcc-c++
-BuildRequires:  gcc-fortran
-BuildRequires:  hipcc
-BuildRequires:  lld
-BuildRequires:  llvm
-BuildRequires:  rocm-cmake
-BuildRequires:  rocm-llvm-macros
-
-Provides:       hipblas = %{version}-%{release}
-
-%description
-hipBLAS is a Basic Linear Algebra Subprograms (BLAS) marshalling
-library, with multiple supported backends. It sits between the
-application and a 'worker' BLAS library, marshalling inputs into
-the backend library and marshalling results back to the
-application. hipBLAS exports an interface that does not require
-the client to change, regardless of the chosen backend. Currently,
-hipBLAS supports rocBLAS and cuBLAS as backends.
-
-%package devel
-Summary:        Libraries and headers for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-Requires:       cmake(hipblas-common)
-
-%description devel
-%{summary}
-
-%if %{with test}
-%package test
-Summary:        Tests for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description test
-%{summary}
-%endif
-
-%prep -a
-# This is a tarball, no .git to query
-sed -i -e 's@find_package(Git REQUIRED)@#find_package(Git REQUIRED)@' library/CMakeLists.txt
-
-%build -a
-rm -f %{buildroot}%{_prefix}/share/doc/hipblas/LICENSE.md
-
-%files
-%license LICENSE.md
-%doc README.md
-%{_libdir}/libhipblas.so.3{,.*}
-
-%files devel
-%{_includedir}/hipblas/
-%{_libdir}/libhipblas.so
-%{_libdir}/cmake/hipblas/
-
-%if %{with test}
-%files test
-%{_bindir}/hipblas*
-%endif
-
-%changelog
-%{?autochangelog}
@@ -0,0 +1,25 @@
+From 43c4a61c5d8836a16feb8e53c72f255790523ff3 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 3 Nov 2025 06:11:40 -0800
+Subject: [PATCH] hipblaslt find origami package
+
+---
+ CMakeLists.txt | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index dbccca92c84f..d02df50540c2 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -218,7 +218,7 @@ if(HIPBLASLT_ENABLE_MSGPACK)
+     endif()
+ endif()
+ 
+-add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/../../shared/origami" origami)
+find_package(origami CONFIG REQUIRED)
+ add_subdirectory(tensilelite)
+ 
+ if(HIPBLASLT_ENABLE_HOST)
+-- 
+2.52.0
+
@@ -0,0 +1,57 @@
+From 72521fcca77c010b9c8b9ce91cde925164502d6f Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Thu, 25 Sep 2025 13:02:55 -0700
+Subject: [PATCH] hipblaslt tensilelite remove yappi dependency
+
+Signed-off-by: Tom Rix <Tom.Rix@amd.com>
+---
+ tensilelite/Tensile/TensileCreateLibrary/Run.py | 15 ---------------
+ tensilelite/requirements.txt                    |  2 +-
+ 2 files changed, 1 insertion(+), 16 deletions(-)
+
+diff --git a/tensilelite/Tensile/TensileCreateLibrary/Run.py b/tensilelite/Tensile/TensileCreateLibrary/Run.py
+index 835ed9c01916..a02705f6554a 100644
+--- a/tensilelite/Tensile/TensileCreateLibrary/Run.py
+++ b/tensilelite/Tensile/TensileCreateLibrary/Run.py
+@@ -231,12 +231,6 @@ def writeSolutionsAndKernels(
+     generateSourcesAndExit=False,
+     compress=True,
+ ):
+-    if globalParameters["PythonProfile"]:
+-        globalParameters["CpuThreads"] = 0
+-        printWarning("Python profiling is enabled. CpuThreads set to 0.")
+-        import yappi
+-        yappi.start()
+-
+     codeObjectFiles = []
+ 
+     outputPath = Path(outputPath)
+@@ -299,15 +293,6 @@ def writeSolutionsAndKernels(
+     writeHelpers(outputPath, kernelHelperObjs, KERNEL_HELPER_FILENAME_CPP, KERNEL_HELPER_FILENAME_H)
+     srcKernelFile = Path(outputPath) / "Kernels.cpp"
+ 
+-    if globalParameters["PythonProfile"]:
+-        yappi.stop()
+-        yappi.get_func_stats().save("yappi_results.profile", type="callgrind")
+-        with open("yappi_results.txt", "w") as f:
+-            yappi.get_func_stats().print_all(out=f)
+-        if globalParameters["CpuThreads"] != 0:
+-            with open("yappi_thread_stats.txt", "w") as f:
+-                yappi.get_thread_stats().print_all(out=f)
+-
+     if not generateSourcesAndExit:
+         codeObjectFiles += buildAssemblyCodeObjectFiles(
+             asmToolchain.linker,
+diff --git a/tensilelite/requirements.txt b/tensilelite/requirements.txt
+index 60c4c1144537..e87db8445411 100644
+--- a/tensilelite/requirements.txt
+++ b/tensilelite/requirements.txt
+@@ -7,4 +7,4 @@ joblib>=1.1.1; python_version < '3.8'
+ simplejson
+ ujson
+ orjson
+-yappi
+
+-- 
+2.52.0
+
@@ -0,0 +1,41 @@
+From 1ac117ac0591a0f1bb67c34f537354c21412b2d8 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Sat, 1 Nov 2025 09:43:58 -0700
+Subject: [PATCH] hipblaslt tensilelite use fedora paths
+
+---
+ tensilelite/Tensile/Common/GlobalParameters.py | 2 +-
+ tensilelite/Tensile/Toolchain/Validators.py    | 4 ++--
+ 2 files changed, 3 insertions(+), 3 deletions(-)
+
+diff --git a/tensilelite/Tensile/Common/GlobalParameters.py b/tensilelite/Tensile/Common/GlobalParameters.py
+index 567188da59bd..1f8037c183a6 100644
+--- a/tensilelite/Tensile/Common/GlobalParameters.py
+++ b/tensilelite/Tensile/Common/GlobalParameters.py
+@@ -538,7 +538,7 @@ def assignGlobalParameters(config, isaInfoMap: Dict[IsaVersion, IsaInfo]):
+         else:
+             print2(" %24s: %8s (unspecified)" % (key, defaultValue))
+ 
+-    globalParameters["ROCmPath"] = "/opt/rocm"
+    globalParameters["ROCmPath"] = "/usr"
+     if "ROCM_PATH" in os.environ:
+         globalParameters["ROCmPath"] = os.environ.get("ROCM_PATH")
+     if "TENSILE_ROCM_PATH" in os.environ:
+diff --git a/tensilelite/Tensile/Toolchain/Validators.py b/tensilelite/Tensile/Toolchain/Validators.py
+index fd5dab5324c0..3ce024d31f52 100644
+--- a/tensilelite/Tensile/Toolchain/Validators.py
+++ b/tensilelite/Tensile/Toolchain/Validators.py
+@@ -30,8 +30,8 @@ from typing import List, NamedTuple, Union
+ 
+ from Tensile.Common.Utilities import isRhel8
+ 
+-DEFAULT_ROCM_BIN_PATH_POSIX = Path("/opt/rocm/bin")
+-DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/opt/rocm/lib/llvm/bin")
+DEFAULT_ROCM_BIN_PATH_POSIX = Path("/usr/bin")
+DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/usr/lib64/rocm/llvm/bin")
+ DEFAULT_ROCM_BIN_PATH_WINDOWS = Path("C:/Program Files/AMD/ROCm")
+ 
+ 
+-- 
+2.52.0
+
@@ -0,0 +1,31 @@
+From 9aa4664e02e27b50083be08e5b495cbef02d6f08 Mon Sep 17 00:00:00 2001
+From: Sakura286 <sakura286@outlook.com>
+Date: Mon, 8 Jun 2026 09:00:08 +0800
+Subject: [PATCH] hipblaslt tensilelite use system nanobind
+
+---
+ tensilelite/rocisa/CMakeLists.txt | 8 +-------
+ 1 file changed, 1 insertion(+), 7 deletions(-)
+
+diff --git a/tensilelite/rocisa/CMakeLists.txt b/tensilelite/rocisa/CMakeLists.txt
+index 3918f18..9c6fcd3 100644
+--- a/tensilelite/rocisa/CMakeLists.txt
+++ b/tensilelite/rocisa/CMakeLists.txt
+@@ -17,13 +17,7 @@ target_include_directories(rocisa-cpp
+ )
+ 
+ if(HIPBLASLT_BUNDLE_PYTHON_DEPS)
+-    include(FetchContent)
+-    FetchContent_Declare(
+-      nanobind
+-      GIT_REPOSITORY https://github.com/wjakob/nanobind.git
+-      GIT_TAG        9b3afa9dbdc23641daf26fadef7743e7127ff92f # v2.6.1
+-    )
+-    FetchContent_MakeAvailable(nanobind)
+    find_package(nanobind CONFIG REQUIRED)
+ 
+     set(ROCISAINST_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/instruction.cpp"
+                           "${CMAKE_CURRENT_SOURCE_DIR}/rocisa/src/instruction/common.cpp"
+-- 
+2.53.0
+
@@ -0,0 +1,197 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+%global toolchain clang
+
+%bcond build_test 0
+%if %{with build_test}
+%global cmake_test ON
+%else
+%global cmake_test OFF
+%endif
+
+%global tensile_version 4.33.0
+# The upstream hipBLASTLt project has a hard fork of the python-tensile package
+# The rocBLAS uses.  The two versions are incompatible.  It appears that the
+# fork happened around version 4.33.0.  Unfortunately hipBLASLt can no longer be
+# build without using this fork.
+# https://github.com/ROCm/hipBLASLt/issues/535
+# The problem with the fork has been raised here.
+# https://github.com/ROCm/hipBLASLt/issues/908
+
+%global tensile_verbose 1
+
+Name:           hipblaslt
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm general matrix operations beyond BLAS
+License:        MIT AND BSD-3-Clause
+URL:            https://github.com/ROCm/rocm-libraries
+#!RemoteAsset:  sha256:05d73038b1b4f66f3df4eb595b7cb0c8935f7aa18d0e07dbe5cc740a4b691898
+Source0:        %{url}/releases/download/rocm-%{version}/%{name}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{cmake_test}
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DHIPBLASLT_ENABLE_CLIENT=%{cmake_test}
+BuildOption(conf):  -DHIPBLASLT_ENABLE_MARKER=OFF
+BuildOption(conf):  -DHIPBLASLT_ENABLE_OPENMP=OFF
+BuildOption(conf):  -DHIPBLASLT_ENABLE_ROCROLLER=OFF
+BuildOption(conf):  -DHIPBLASLT_ENABLE_SAMPLES=OFF
+BuildOption(conf):  -DTensile_LIBRARY_FORMAT=msgpack
+BuildOption(conf):  -DTensile_VERBOSE=%{tensile_verbose}
+BuildOption(conf):  -DVIRTUALENV_BIN_DIR=%{_bindir}
+BuildOption(conf):  -Dnanobind_ROOT=%(python3 -m nanobind --cmake_dir)
+BuildOption(conf):  -G Ninja
+
+# yappi is used in tensilelite to generate profiling data, we are not using that in the build
+Patch0:         0001-hipblaslt-tensilelite-remove-yappi-dependency.patch
+# Patch from Fedora, change hard coded vendor paths
+Patch1:         0001-hipblaslt-tensilelite-use-system-paths.patch
+# https://github.com/ROCm/rocm-libraries/issues/2422
+Patch2:         0001-hipblaslt-find-origami-package.patch
+# use the distribution-provided nanobind instead of fetching/bundling it
+Patch3:         2001-hipblaslt-tensilelite-use-system-nanobind.patch
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hipblas)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(msgpack)
+BuildRequires:  cmake(origami)
+BuildRequires:  cmake(rocblas)
+BuildRequires:  cmake(rocm_smi)
+BuildRequires:  compiler-rt
+BuildRequires:  gcc-fortran
+BuildRequires:  hipcc
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(libzstd)
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  pkgconfig(zlib)
+# https://github.com/ROCm/hipBLASLt/issues/1734
+BuildRequires:  python3dist(msgpack)
+# nanobind is used to build the rocisa native module (build-time only)
+BuildRequires:  python3dist(nanobind)
+BuildRequires:  python3dist(setuptools)
+BuildRequires:  python3dist(pyyaml)
+BuildRequires:  python3dist(joblib)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+BuildRequires:  rocminfo
+
+%if %{with build_test}
+BuildRequires:  cmake(openblas)
+BuildRequires:  cmake(GMock)
+BuildRequires:  cmake(GTest)
+%endif
+
+%description
+hipBLASLt is a library that provides general matrix-matrix
+operations. It has a flexible API that extends functionalities
+beyond a traditional BLAS library, such as adding flexibility
+to matrix data layouts, input types, compute types, and
+algorithmic implementations and heuristics.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%if %{with build_test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep -a
+# Use PATH to find where TensileGetPath and other tensile bins are
+sed -i -e 's@${Tensile_PREFIX}/bin/TensileGetPath@TensileGetPath@g'            tensilelite/Tensile/cmake/TensileConfig.cmake
+
+# defer to cmdline
+sed -i -e 's@set(CMAKE_INSTALL_LIBDIR@#set(CMAKE_INSTALL_LIBDIR@' CMakeLists.txt
+
+# Do not use virtualenv_install
+sed -i -e 's@virtualenv_install@#virtualenv_install@'                          CMakeLists.txt
+
+# Disable trying to download rocm-cmake
+sed -i -e 's@if(NOT ROCmCMakeBuildTools_FOUND)@if(FALSE)@' cmake/dependencies.cmake
+
+# HIPBLASLT_ENABLE_OPENMP is OFF yet it is still being used
+# https://github.com/ROCm/rocm-libraries/issues/3201
+sed -i -e '/OpenMP::OpenMP_CXX/d' clients/CMakeLists.txt
+sed -i -e '/omp/d'                clients/common/src/blis_interface.cpp
+sed -i -e '/#include <omp.h>/d'   clients/common/include/testing_matmul.hpp
+sed -i -e '/#include <omp.h>/d'   clients/common/include/hipblaslt_init.hpp
+sed -i -e '/#include <omp.h>/d'   clients/common/src/cblas_interface.cpp
+
+# We are building from a tarball, not a git repo
+sed -i -e 's@find_package(Git REQUIRED)@#find_package(Git REQUIRED)@' cmake/dependencies.cmake
+
+# Forcefully replace all mentions of 'amdclang' with 'clang' in the Tensile Python files
+find tensilelite -type f -name "*.py" -exec sed -i 's/amdclang++/clang++/g; s/amdclang/clang/g' {} +
+
+%build -p
+# Do a manual install instead of cmake's virtualenv
+cd tensilelite
+TL=$PWD
+
+python3 setup.py install --root $TL
+cd ..
+
+# Should not have to do this
+CLANG_PATH=`hipconfig --hipclangpath`
+ROCM_CLANG=${CLANG_PATH}/clang
+RESOURCE_DIR=`${ROCM_CLANG} -print-resource-dir`
+export DEVICE_LIB_PATH=${RESOURCE_DIR}/amdgcn/bitcode
+export TENSILE_ROCM_ASSEMBLER_PATH=${CLANG_PATH}/clang++
+export TENSILE_ROCM_OFFLOAD_BUNDLER_PATH=${CLANG_PATH}/clang-offload-bundler
+
+# Look for the just built tensilelite
+export PATH=${TL}/%{_bindir}:$PATH
+export PYTHONPATH=${TL}%{python3_sitelib}:$PYTHONPATH
+export Tensile_DIR=${TL}%{python3_sitelib}/Tensile
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hipblaslt/LICENSE.md
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhipblaslt.so.*
+%{_libdir}/hipblaslt/
+
+%files devel
+%{_includedir}/hipblaslt/
+%{_includedir}/hipblaslt-export.h
+%{_includedir}/hipblaslt-version.h
+%{_libdir}/cmake/hipblaslt/
+%{_libdir}/libhipblaslt.so
+
+%if %{with build_test}
+%files test
+%{_bindir}/hipblaslt*
+%{_bindir}/sequence.yaml
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,84 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%bcond test 0
+%if %{with test}
+%global build_test ON
+%else
+%global build_test OFF
+%endif
+
+%global toolchain clang
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+Name:           hipcub
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm port of CUDA CUB (header-only)
+License:        BSD-3-Clause AND MIT
+Url:            https://github.com/ROCm/rocm-libraries
+#!RemoteAsset:  sha256:6dadbb7689c7906493ec42f56792d9557f0293670a86059c9c188851f399647b
+Source:         %{url}/releases/download/rocm-%{version}/hipcub.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=%{build_test}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocprim)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-llvm-macros
+
+%description
+hipCUB is a thin header-only wrapper library on top of rocPRIM which enables
+developers to render portable HIP code. Existing CUDA CUB source code can
+be recompiled in HIP using hipCUB.
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep -a
+# Fix cmake install lib directory
+sed -i -e 's/ROCM_INSTALL_LIBDIR lib/ROCM_INSTALL_LIBDIR %{_lib}/' \
+    cmake/ROCMExportTargetsHeaderOnly.cmake
+
+%install -a
+rm -f %{buildroot}/%{_datadir}/doc/hipcub/LICENSE.txt
+
+%files
+%doc README.md
+%license LICENSE.txt
+%{_includedir}/hipcub/
+%{_libdir}/cmake/hipcub/
+
+%if %{with test}
+%files test
+%{_bindir}/test_*
+%{_bindir}/hipcub/
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,24 @@
+From cfa8e85698486f791008fcade4ec2dff8ddd99d9 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Fri, 31 Oct 2025 09:10:07 -0700
+Subject: [PATCH] hipfft hipfftw soversion
+
+---
+ library/CMakeLists.txt | 1 +
+ 1 file changed, 1 insertion(+)
+
+diff --git a/library/CMakeLists.txt b/library/CMakeLists.txt
+index 8c97cc4eeb97..97cacb3451d3 100644
+--- a/library/CMakeLists.txt
+++ b/library/CMakeLists.txt
+@@ -164,6 +164,7 @@ endif()
+ # nvcc can not recognize shared library file name with suffix other than *.so when linking.
+ if (NOT BUILD_WITH_COMPILER STREQUAL "HIP-NVCC")
+   rocm_set_soversion(hipfft ${hipfft_SOVERSION})
+  rocm_set_soversion(hipfftw ${hipfft_SOVERSION})
+ endif()
+ 
+ # Generate export header
+-- 
+2.51.0
+
@@ -0,0 +1,105 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%bcond test 1
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           hipfft
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm FFT marshalling library
+Url:            https://github.com/ROCm/rocm-libraries
+VCS:            git:https://github.com/ROCm/hipFFT.git
+License:        MIT
+#!RemoteAsset:  sha256:f6f0352b5f9ffe53c88cea5fa40572eef0c0c1e2e50dce6f85d2c68e47afc63e
+Source:         %{url}/releases/download/rocm-%{version}/hipfft.tar.gz
+Patch1:         0001-hipfft-hipfftw-soversion.patch
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=ON
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS_OPENMP=OFF
+
+BuildRequires:  boost-devel
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(GTest)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hiprand)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocfft)
+BuildRequires:  cmake(rocrand)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(fftw3)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+hipFFT is a FFT marshalling library. Currently, hipFFT supports either
+rocFFT or cuFFT as backends. hipFFT exports an interface that does not
+require the client to change, regardless of the chosen backend.
+
+%package        devel
+Summary:        The hipFFT development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(rocfft)
+
+%description    devel
+The hipFFT development package.
+
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+
+%prep -a
+# CMake Error at clients/tests/CMakeLists.txt:87 (find_package):
+#   No "FindHIP.cmake" found in CMAKE_MODULE_PATH.
+# Remove MODULE
+sed -i -e 's@find_package( HIP MODULE REQUIRED )@find_package( HIP REQUIRED )@' \
+    clients/tests/CMakeLists.txt
+
+%install -a
+rm -f %{buildroot}/%{_datadir}/doc/hipfft/LICENSE.md
+
+%if %{with test}
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH
+%endif
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhipfft.so.0{,.*}
+%{_libdir}/libhipfftw.so.0{,.*}
+
+%files devel
+%{_includedir}/hipfft/
+%{_libdir}/cmake/hipfft/
+%{_libdir}/libhipfft.so
+%{_libdir}/libhipfftw.so
+
+%files test
+%{_bindir}/hipfft-test
+
+%changelog
+%autochangelog
@@ -0,0 +1,102 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# HIP error 100: no ROCm-capable device is detected
+# hipRAND needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU, so make it optional.
+%bcond run_test 0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           hiprand
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        HIP random number generator
+License:        MIT AND BSD-3-Clause
+Url:            https://github.com/ROCm/rocm-libraries
+#!RemoteAsset:  sha256:41e4053a3c16ea4bdc6e94fff428d8ffe7279e9cfa7ec142afc50169aae2c1f8
+Source:         %{url}/releases/download/rocm-%{version}/hiprand.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=ON
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(GTest)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocrand)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+hipRAND is a RAND marshalling library, with multiple supported backends. It
+sits between the application and the backend RAND library, marshalling inputs
+into the backend and results back to the application. hipRAND exports an
+interface that does not require the client to change, regardless of the chosen
+backend. Currently, hipRAND supports either rocRAND or cuRAND.
+
+%package        devel
+Summary:        The hipRAND development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(rocrand)
+
+%description    devel
+The hipRAND development package.
+
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+
+%prep -a
+# Remove RPATH
+sed -i '/INSTALL_RPATH/d' CMakeLists.txt
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hiprand/LICENSE.md
+rm -f %{buildroot}%{_bindir}/hipRAND/CTestTestfile.cmake
+
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH
+
+%if %{without run_test}
+%check
+%endif
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhiprand.so.1{,.*}
+
+%files devel
+%{_includedir}/hiprand/
+%{_libdir}/cmake/hiprand/
+%{_libdir}/libhiprand.so
+
+%files test
+%{_bindir}/test*
+
+%changelog
+%autochangelog
@@ -0,0 +1,24 @@
+From 2d228ee883a538d67af24e944a2a24a95d779345 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Sun, 21 Sep 2025 08:37:33 -0700
+Subject: [PATCH] hipsolver so version fortran bindings
+
+---
+ library/src/CMakeLists.txt | 1 +
+ 1 file changed, 1 insertion(+)
+
+diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
+index a4e8649cdcae..69437d603d7a 100644
+--- a/library/src/CMakeLists.txt
+++ b/library/src/CMakeLists.txt
+@@ -75,6 +75,7 @@ set(hipsolver_f90_source
+ if(BUILD_FORTRAN_BINDINGS)
+   # Create hipSOLVER Fortran module
+   add_library(hipsolver_fortran ${hipsolver_f90_source})
+  rocm_set_soversion(hipsolver_fortran ${hipsolver_SOVERSION})
+   rocm_install(TARGETS hipsolver_fortran)
+ endif()
+ 
+-- 
+2.51.0
+
@@ -0,0 +1,124 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# TODO: hipSOLVER need lapack to build test/benchmark/sample
+# But openblas on openRuyi does not provide this
+%bcond build_test 0
+%if %{with build_test}
+%global cmake_test ON
+%else
+%global cmake_test OFF
+%endif
+
+# hipSOLVER needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU.
+%bcond run_test 0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+# Fortran is only used in testing
+# clang and gfortran fedora toolchain args do not mix
+%global build_fflags %{nil}
+
+Name:           hipsolver
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm SOLVER marshalling library (LAPACK)
+License:        MIT
+Url:            https://github.com/ROCm/hipSOLVER
+#!RemoteAsset:  sha256:bd664e3cd43bfcc7e94d5a387c27262c4b218d6d2e71e086992b174349dd1c10
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{cmake_test}
+BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=%{cmake_test}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocblas)
+BuildRequires:  cmake(rocsolver)
+BuildRequires:  cmake(rocsparse)
+BuildRequires:  compiler-rt
+BuildRequires:  gcc-fortran
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+%if %{with build_test}
+BuildRequires:  cmake(GTest)
+BuildRequires:  cmake(hipsparse)
+BuildRequires:  pkgconfig(openblas)
+%endif
+
+%description
+hipSOLVER is a LAPACK marshalling library, with multiple supported backends.
+It sits between the application and a "worker" SOLVER library, marshalling
+inputs into the backend library and results back to the application. hipSOLVER
+exports an interface that does not require the client to change, regardless of
+the chosen backend.
+
+%package        devel
+Summary:        The hipSOLVER development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(rocblas)
+Requires:       cmake(rocsolver)
+Requires:       cmake(rocsparse)
+
+%description    devel
+The hipSOLVER development package.
+
+%if %{with build_test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hipsolver/LICENSE.md
+
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH
+
+%if %{without test}
+%check
+%endif
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhipsolver.so.1{,.*}
+%{_libdir}/libhipsolver_fortran.so.1{,.*}
+
+%files devel
+%{_includedir}/hipsolver/
+%{_libdir}/libhipsolver.so
+%{_libdir}/libhipsolver_fortran.so
+%{_libdir}/cmake/hipsolver/
+
+%if %{with build_test}
+%files test
+%{_datadir}/hipsolver/
+%{_bindir}/hipsolver*
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,128 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# hipSPARSE need to download about 19 testing matrix
+# It is verbose to add them to SOURCE and %%prep section
+%bcond build_test 0
+%if %{with build_test}
+%global cmake_test ON
+%else
+%global cmake_test OFF
+%endif
+
+# hipSPARSE needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU.
+%bcond run_test 0
+
+# This ROCm package is built with clang by default
+%global toolchain clang
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+Name:           hipsparse
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm SPARSE marshalling library
+License:        MIT
+Url:            https://github.com/ROCm/hipSPARSE
+#!RemoteAsset:  sha256:b001834d8e65c3878d1a69d08803d5b6ce4fe623e78099fe51cb146d0ffa10e7
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_CLIENTS_SAMPLES=OFF
+BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=ON
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{cmake_test}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+%if %{with build_test}
+BuildRequires:  cmake(GTest)
+%endif
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocprim)
+BuildRequires:  cmake(rocsparse)
+BuildRequires:  compiler-rt
+BuildRequires:  gcc-fortran
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+hipSPARSE is a SPARSE marshalling library with multiple
+supported backends. It sits between your application and
+a 'worker' SPARSE library, where it marshals inputs to
+the backend library and marshals results to your
+application. hipSPARSE exports an interface that doesn't
+require the client to change, regardless of the chosen
+backend. Currently, hipSPARSE supports rocSPARSE and
+cuSPARSE backends.
+
+%package        benchmark
+Summary:        Benchmark for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    benchmark
+%{summary}
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%if %{with build_test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hipsparse/LICENSE.md
+
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH
+
+%if %{without run_test}
+%check
+%endif
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhipsparse.so.4{,.*}
+
+%files benchmark
+%{_bindir}/hipsparse-bench
+
+%files devel
+%{_includedir}/hipsparse/
+%{_libdir}/cmake/hipsparse/
+%{_libdir}/libhipsparse.so
+
+%if %{with build_test}
+%files test
+%{_bindir}/hipsparse*
+%{_datadir}/hipsparse/
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,25 @@
+From 43c4a61c5d8836a16feb8e53c72f255790523ff3 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 3 Nov 2025 06:11:40 -0800
+Subject: [PATCH] hipblaslt find origami package
+
+---
+ CMakeLists.txt | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index dbccca92c84f..d02df50540c2 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -218,7 +218,7 @@ if(HIPBLASLT_ENABLE_MSGPACK)
+     endif()
+ endif()
+ 
+-add_subdirectory("${CMAKE_CURRENT_SOURCE_DIR}/../../shared/origami" origami)
+find_package(origami CONFIG REQUIRED)
+ add_subdirectory(tensilelite)
+ 
+ if(HIPBLASLT_ENABLE_HOST)
+-- 
+2.52.0
+
@@ -0,0 +1,57 @@
+From c20b846b6d594464eccf865045ef0ef10384f407 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Thu, 25 Sep 2025 13:02:55 -0700
+Subject: [PATCH] hipblaslt tensilelite remove yappi dependency
+
+Signed-off-by: Tom Rix <Tom.Rix@amd.com>
+---
+ tensilelite/Tensile/TensileCreateLibrary/Run.py | 15 ---------------
+ tensilelite/requirements.txt                    |  2 +-
+ 2 files changed, 1 insertion(+), 16 deletions(-)
+
+diff --git a/tensilelite/Tensile/TensileCreateLibrary/Run.py b/tensilelite/Tensile/TensileCreateLibrary/Run.py
+index f0bbe8acd127..fc076e6935e8 100644
+--- a/tensilelite/Tensile/TensileCreateLibrary/Run.py
+++ b/tensilelite/Tensile/TensileCreateLibrary/Run.py
+@@ -231,12 +231,6 @@ def writeSolutionsAndKernels(
+     generateSourcesAndExit=False,
+     compress=True,
+ ):
+-    if globalParameters["PythonProfile"]:
+-        globalParameters["CpuThreads"] = 0
+-        printWarning("Python profiling is enabled. CpuThreads set to 0.")
+-        import yappi
+-        yappi.start()
+-
+     codeObjectFiles = []
+ 
+     outputPath = Path(outputPath)
+@@ -299,15 +293,6 @@ def writeSolutionsAndKernels(
+     writeHelpers(outputPath, kernelHelperObjs, KERNEL_HELPER_FILENAME_CPP, KERNEL_HELPER_FILENAME_H)
+     srcKernelFile = Path(outputPath) / "Kernels.cpp"
+ 
+-    if globalParameters["PythonProfile"]:
+-        yappi.stop()
+-        yappi.get_func_stats().save("yappi_results.profile", type="callgrind")
+-        with open("yappi_results.txt", "w") as f:
+-            yappi.get_func_stats().print_all(out=f)
+-        if globalParameters["CpuThreads"] != 0:
+-            with open("yappi_thread_stats.txt", "w") as f:
+-                yappi.get_thread_stats().print_all(out=f)
+-
+     if not generateSourcesAndExit:
+         codeObjectFiles += buildAssemblyCodeObjectFiles(
+             asmToolchain.linker,
+diff --git a/tensilelite/requirements.txt b/tensilelite/requirements.txt
+index 60c4c1144537..e87db8445411 100644
+--- a/tensilelite/requirements.txt
+++ b/tensilelite/requirements.txt
+@@ -7,4 +7,4 @@ joblib>=1.1.1; python_version < '3.8'
+ simplejson
+ ujson
+ orjson
+-yappi
+
+-- 
+2.51.0
+
@@ -0,0 +1,16 @@
+--- ./hipBLASLt/tensilelite/Tensile/Toolchain/Validators.py	2025-11-14 05:43:51
+++ ./hipBLASLt/tensilelite/Tensile/Toolchain/Validators.py.mod	2026-03-04 16:10:09
+@@ -114,11 +114,11 @@
+
+
+ class ToolchainDefaults(NamedTuple):
+-    CXX_COMPILER = osSelect(linux="amdclang++", windows="clang++.exe")
+-    C_COMPILER = osSelect(linux="amdclang", windows="clang.exe")
+    CXX_COMPILER = osSelect(linux="clang++", windows="clang++.exe")
+    C_COMPILER = osSelect(linux="clang", windows="clang.exe")
+     OFFLOAD_BUNDLER = osSelect(linux="clang-offload-bundler", windows="clang-offload-bundler.exe")
+     DEVICE_ENUMERATOR = osSelect(linux="rocm_agent_enumerator" if isRhel8() else "amdgpu-arch", windows="hipinfo")
+-    ASSEMBLER = osSelect(linux="amdclang++", windows="clang++.exe")
+    ASSEMBLER = osSelect(linux="clang++", windows="clang++.exe")
+     HIP_CONFIG = osSelect(linux="hipconfig", windows="hipconfig.exe")
+
@@ -0,0 +1,41 @@
+From 1ac117ac0591a0f1bb67c34f537354c21412b2d8 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Sat, 1 Nov 2025 09:43:58 -0700
+Subject: [PATCH] hipblaslt tensilelite use fedora paths
+
+---
+ tensilelite/Tensile/Common/GlobalParameters.py | 2 +-
+ tensilelite/Tensile/Toolchain/Validators.py    | 4 ++--
+ 2 files changed, 3 insertions(+), 3 deletions(-)
+
+diff --git a/tensilelite/Tensile/Common/GlobalParameters.py b/tensilelite/Tensile/Common/GlobalParameters.py
+index 567188da59bd..1f8037c183a6 100644
+--- a/tensilelite/Tensile/Common/GlobalParameters.py
+++ b/tensilelite/Tensile/Common/GlobalParameters.py
+@@ -538,7 +538,7 @@ def assignGlobalParameters(config, isaInfoMap: Dict[IsaVersion, IsaInfo]):
+         else:
+             print2(" %24s: %8s (unspecified)" % (key, defaultValue))
+ 
+-    globalParameters["ROCmPath"] = "/opt/rocm"
+    globalParameters["ROCmPath"] = "/usr"
+     if "ROCM_PATH" in os.environ:
+         globalParameters["ROCmPath"] = os.environ.get("ROCM_PATH")
+     if "TENSILE_ROCM_PATH" in os.environ:
+diff --git a/tensilelite/Tensile/Toolchain/Validators.py b/tensilelite/Tensile/Toolchain/Validators.py
+index fd5dab5324c0..3ce024d31f52 100644
+--- a/tensilelite/Tensile/Toolchain/Validators.py
+++ b/tensilelite/Tensile/Toolchain/Validators.py
+@@ -30,8 +30,8 @@ from typing import List, NamedTuple, Union
+ 
+ from Tensile.Common.Utilities import isRhel8
+ 
+-DEFAULT_ROCM_BIN_PATH_POSIX = Path("/opt/rocm/bin")
+-DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/opt/rocm/lib/llvm/bin")
+DEFAULT_ROCM_BIN_PATH_POSIX = Path("/usr/bin")
+DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/usr/lib64/rocm/llvm/bin")
+ DEFAULT_ROCM_BIN_PATH_WINDOWS = Path("C:/Program Files/AMD/ROCm")
+ 
+ 
+-- 
+2.52.0
+
@@ -0,0 +1,205 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+%global toolchain clang
+
+%global tensile_version 4.33.0
+%global tensile_verbose 1
+
+%bcond build_test 0
+%if %{with build_test}
+%global cmake_test ON
+%else
+%global cmake_test OFF
+%endif
+
+Name:           hipsparselt
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        A SPARSE marshaling library
+License:        MIT
+URL:            https://github.com/ROCm/rocm-libraries
+#!RemoteAsset:  sha256:7672d1ac94d2694999b6937d19f5e92e67fb844eea394b4e8525c531fd1acd8c
+Source0:        %{url}/releases/download/rocm-%{version}/%{name}.tar.gz
+#!RemoteAsset:  sha256:05d73038b1b4f66f3df4eb595b7cb0c8935f7aa18d0e07dbe5cc740a4b691898
+Source1:        %{url}/releases/download/rocm-%{version}/hipblaslt.tar.gz
+# Patches for hipBLASLt's tensilelite (applied during prep inside hipBLASLt/)
+Source2:        0001-hipblaslt-tensilelite-remove-yappi-dependency.patch
+Source3:        0001-hipblaslt-tensilelite-use-system-paths.patch
+Source4:        0001-hipblaslt-find-origami-package.patch
+BuildSystem:    cmake
+
+BuildOption(conf):  -DBLAS_INCLUDE_DIR=%{_includedir}/flexiblas
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{cmake_test}
+BuildOption(conf):  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
+BuildOption(conf):  -DBUILD_VERBOSE=ON
+BuildOption(conf):  -DCMAKE_Fortran_COMPILER=gcc-fortran
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DTensile_COMPILER=clang++
+BuildOption(conf):  -DTensile_LIBRARY_FORMAT=msgpack
+BuildOption(conf):  -DTensile_VERBOSE=%{tensile_verbose}
+BuildOption(conf):  -DVIRTUALENV_BIN_DIR=%{_bindir}
+BuildOption(conf):  -Dnanobind_ROOT=%(python3 -m nanobind --cmake_dir)
+BuildOption(conf):  -G Ninja
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hipsparse)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(origami)
+BuildRequires:  cmake(rocm_smi)
+BuildRequires:  cmake(rocsparse)
+BuildRequires:  compiler-rt
+BuildRequires:  gcc-fortran
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(libzstd)
+BuildRequires:  pkgconfig(msgpack)
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  pkgconfig(zlib)
+BuildRequires:  python3dist(joblib)
+BuildRequires:  python3dist(msgpack)
+# nanobind is used to build the rocisa native module (build-time only)
+BuildRequires:  python3dist(nanobind)
+BuildRequires:  python3dist(pyyaml)
+BuildRequires:  python3dist(setuptools)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocminfo
+BuildRequires:  rocm-llvm-macros
+BuildRequires:  roctracer-devel
+
+%if %{with build_test}
+BuildRequires:  chrpath
+BuildRequires:  pkgconfig(openblas)
+BuildRequires:  pkgconfig(gtest)
+BuildRequires:  pkgconfig(gmock)
+%endif
+
+%description
+hipSPARSELt is a SPARSE marshaling library that provides general sparse
+matrix-matrix multiplication using structured sparsity. It offers a flexible
+API and supports multiple backends.
+
+%package        devel
+Summary:        The hipSPARSELt development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+The hipSPARSELt development package.
+
+%if %{with build_test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep
+%autosetup -p1 -n %{name}
+
+tar xf %{SOURCE1}
+cd hipblaslt
+
+patch -p1 < %{SOURCE2}
+patch -p1 < %{SOURCE3}
+patch -p1 < %{SOURCE4}
+
+# Use PATH to find where TensileGetPath and other tensile bins are
+sed -i -e 's@${Tensile_PREFIX}/bin/TensileGetPath@TensileGetPath@g' \
+    tensilelite/Tensile/cmake/TensileConfig.cmake
+
+# Make sure hip/hip_runtime.h is found
+sed -i -e 's@-x hip @-I%{_includedir} -x hip @' device-library/matrix-transform/CMakeLists.txt
+sed -i -e 's@"-D__HIP_HCC_COMPAT_MODE__=1"@"-D__HIP_HCC_COMPAT_MODE__=1","-I%{_includedir}"@' \
+    tensilelite/Tensile/Toolchain/Component.py
+
+# Use the distribution-provided nanobind instead of fetching/bundling it
+sed -i -e 's@FetchContent_MakeAvailable(nanobind)@find_package(nanobind CONFIG REQUIRED)@' \
+    tensilelite/rocisa/CMakeLists.txt
+
+# disable openmp in hipBLASLt
+sed -i -e 's@option(HIPBLASLT_ENABLE_OPENMP "Use OpenMP to improve performance." ON)@option(HIPBLASLT_ENABLE_OPENMP "Use OpenMP to improve performance." OFF)@' CMakeLists.txt
+
+cd ..
+
+# Point hipBLASLt path at the bundled in-source copy (default looks in ../hipblaslt)
+sed -i -e 's@${CMAKE_CURRENT_SOURCE_DIR}/../hipblaslt@${CMAKE_CURRENT_SOURCE_DIR}/hipblaslt@' CMakeLists.txt
+
+# Prevent the virtualenv install from cmake
+sed -i -e 's@virtualenv_install@#virtualenv_install@' CMakeLists.txt
+
+# Unforce the setting of libdir
+sed -i -e 's@set(CMAKE_INSTALL_LIBDIR@#set(CMAKE_INSTALL_LIBDIR@' CMakeLists.txt
+
+# Change looking for cblas to flexiblas
+sed -i -e 's@find_package( cblas REQUIRED CONFIG )@#find_package( cblas REQUIRED CONFIG )@' clients/CMakeLists.txt
+sed -i -e 's@set( BLAS_LIBRARY "blas" )@set( BLAS_LIBRARY "flexiblas" )@' clients/CMakeLists.txt
+sed -i -e 's@lapack cblas@flexiblas@' clients/gtest/CMakeLists.txt
+
+# We are building from a tarball, not a git repo
+sed -i -e 's@find_package(Git REQUIRED)@#find_package(Git REQUIRED)@' hipblaslt/cmake/dependencies.cmake
+sed -i -e 's@find_package(Git REQUIRED)@#find_package(Git REQUIRED)@' cmake/Dependencies.cmake
+
+# Replace all mentions of 'amdclang' with 'clang' in Tensile Python files
+find hipblaslt/tensilelite -type f -name "*.py" -exec sed -i 's/amdclang++/clang++/g; s/amdclang/clang/g' {} +
+
+%build -p
+# Do a manual install of tensilelite instead of cmake's virtualenv, then point
+# Tensile at it for build-time kernel generation (same approach as hipblaslt)
+cd hipblaslt/tensilelite
+TL=$PWD
+python3 setup.py install --root $TL
+cd ../..
+
+export PATH=%{_prefix}/bin:%{rocmllvm_bindir}:$PATH
+CLANG_PATH=`hipconfig --hipclangpath`
+ROCM_CLANG=${CLANG_PATH}/clang
+RESOURCE_DIR=`${ROCM_CLANG} -print-resource-dir`
+export DEVICE_LIB_PATH=${RESOURCE_DIR}/amdgcn/bitcode
+export TENSILE_ROCM_ASSEMBLER_PATH=${CLANG_PATH}/clang++
+export TENSILE_ROCM_OFFLOAD_BUNDLER_PATH=${CLANG_PATH}/clang-offload-bundler
+export PATH=${TL}/%{_bindir}:$PATH
+export PYTHONPATH=${TL}%{python3_sitelib}:$PYTHONPATH
+export Tensile_DIR=${TL}%{python3_sitelib}/Tensile
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hipsparselt/LICENSE.md
+
+# Strip and fix permissions on hsaco kernel files
+%{rocmllvm_bindir}/llvm-strip %{buildroot}%{_libdir}/hipsparselt/library/Kernels*.hsaco
+chmod a+x %{buildroot}%{_libdir}/hipsparselt/library/Kernels*.hsaco
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libhipsparselt.so.*
+%{_libdir}/hipsparselt/
+
+%files devel
+%{_includedir}/hipsparselt/
+%{_libdir}/cmake/hipsparselt/
+%{_libdir}/libhipsparselt.so
+
+%if %{with build_test}
+%files test
+%{_bindir}/hipsparselt*
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,135 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# rocm toolchain uses the hipcc wrapper of clang
+%global toolchain clang
+
+%bcond test 0
+
+Name:           magma
+Version:        2.10.0
+Release:        %autorelease
+Summary:        Matrix Algebra on GPU and Multi-core Architectures
+License:        BSD-3-Clause
+Url:            https://icl.utk.edu/magma/
+VCS:            git:https://github.com/icl-utk-edu/magma.git
+#!RemoteAsset:  sha256:26347adbccbe7a6693d6b3f3c0ab5620037eb3a62b5ef69d05e40289472a82a4
+Source0:        https://github.com/icl-utk-edu/%{name}/archive/v%{version}.tar.gz
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DBLA_VENDOR=OpenBLAS
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DMAGMA_ENABLE_HIP=ON
+BuildOption(conf):  -DUSE_FORTRAN=OFF
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hipblas)
+BuildRequires:  cmake(hipsparse)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(openblas)
+BuildRequires:  compiler-rt
+BuildRequires:  gcc-c++
+BuildRequires:  hipcc
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  python3
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+Matrix Algebra on GPU and Multi-core Architectures (MAGMA) is a collection
+of next-generation linear algebra libraries for heterogeneous computing.
+The MAGMA package supports interfaces for current linear algebra packages
+and standards (e.g., LAPACK and BLAS) to enable computational scientists
+to easily port any linear algebra–reliant software component to
+heterogeneous computing systems. MAGMA enables applications to fully
+exploit the power of current hybrid systems of many-core CPUs and
+multi-GPUs/coprocessors to deliver the fastest possible time to accurate
+solutions within given energy constraints.
+
+MAGMA features LAPACK-compliant routines for multi-core CPUs enhanced with
+NVIDIA or AMD GPUs. MAGMA 2.7.2 now includes more than 400 routines that
+cover one-sided dense matrix factorizations and solvers, two-sided
+factorizations, and eigen/singular-value problem solvers, as well as a
+subset of highly optimized BLAS for GPUs. A MagmaDNN package has been
+added and further enhanced to provide high-performance data analytics,
+including functionalities for machine learning applications that use MAGMA
+as their computational back end. The MAGMA Sparse and MAGMA Batched
+packages have been included since MAGMA 1.6.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%prep -a
+# Add newer gfx targets to Makefile's valid arch whitelist
+# https://bitbucket.org/icl/magma/issues/76/a-few-new-rocm-gpus
+sed -i -e 's@1032 1033@1032 1033 1100 1101 1102 1103 1150 1151 1152 1153 1200 1201@' Makefile
+
+%if %{with test}
+# Remove a test that fails to link (undefined magma_generate_matrix)
+sed -i -e '/testing_zgenerate.cpp/d' testing/Makefile.src
+%else
+# Disable building tests
+sed -i -e 's@include_directories( testing )@#include_directories( testing )@' CMakeLists.txt
+sed -i -e 's@foreach( filename ${testing_all} )@foreach( filename ${no_testing_all} )@' CMakeLists.txt
+sed -i -e 's@add_custom_target( testing DEPENDS ${testing} )@#add_custom_target( testing DEPENDS ${testing} )@' CMakeLists.txt
+sed -i -e 's@foreach( TEST ${sparse_testing_all} )@foreach( TEST ${no_sparse_testing_all} )@' CMakeLists.txt
+sed -i -e 's@add_custom_target( sparse-testing DEPENDS ${sparse-testing} )@#add_custom_target( sparse-testing DEPENDS ${sparse-testing} )@' CMakeLists.txt
+%endif
+
+# Change the bin,lib install locations
+sed -i -e 's@DESTINATION lib@DESTINATION ${CMAKE_INSTALL_LIBDIR}@' CMakeLists.txt
+sed -i -e 's@DESTINATION bin@DESTINATION ${CMAKE_INSTALL_BINDIR}@' CMakeLists.txt
+
+# python to python3, need env to find local bits like magmasubs.py
+sed -i -e 's@env python@env python3@' tools/checklist_run_tests.py
+sed -i -e 's@env python@env python3@' tools/check-style.py
+sed -i -e 's@env python@env python3@' tools/parse-magma.py
+
+# Remove some files we do not need to similify licenses
+# GPL, results for cuda
+rm -rf results/*
+# ICS, Copy of strlcpy - just use strlcpy
+sed -i -e '/strlcpy/d' control/Makefile.src
+sed -i -e '/strlcpy/d' include/magma_auxiliary.h
+sed -i -e 's@magma_strlcpy@strlcpy@' control/trace.cpp
+rm control/strlcpy.cpp
+
+%build -p
+echo "BACKEND = hip"                          > make.inc
+echo "FORT = false"                          >> make.inc
+echo "GPU_TARGET = gfx1100;gfx1200;gfx1201"  >> make.inc
+
+make generate
+%if %{with test}
+%check
+%{_vpath_builddir}/testing/testing_sgemm
+%endif
+
+%files
+%license COPYRIGHT
+%{_libdir}/libmagma.so{,.*}
+%{_libdir}/libmagma_sparse.so{,.*}
+
+%files devel
+%{_includedir}/*.h
+%{_libdir}/pkgconfig/%{name}.pc
+%{_libdir}/libmagma.so
+%{_libdir}/libmagma_sparse.so
+
+%changelog
+%autochangelog
@@ -0,0 +1,147 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# Tests require an AMD GPU; keep the bcond for packagers with hardware.
+%bcond test 0
+
+%global rocm_release 7.1
+%global rocm_patch   1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           miopen
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        AMD's Machine Intelligence Library
+License:        MIT AND BSD-2-Clause AND Apache-2.0
+Url:            https://github.com/ROCm/MIOpen
+#!RemoteAsset:  sha256:98c72a2b5ca541d6c172facdf0f15729207ab52ca9af36c00e2480c5b27c5b99
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+# Adds MIOPEN_PARALLEL_{COMPILE,LINK}_JOBS options to limit Ninja job pools
+# and avoid OOM on memory-constrained build hosts (upstream patch)
+Patch0:         0001-miopen-add-link-and-compile-pools.patch
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBoost_USE_STATIC_LIBS=OFF
+BuildOption(conf):  -DMIOPEN_BUILD_DRIVER=OFF
+BuildOption(conf):  -DMIOPEN_ENABLE_AI_IMMED_MODE_FALLBACK=OFF
+BuildOption(conf):  -DMIOPEN_ENABLE_AI_KERNEL_TUNING=OFF
+%if %{with test}
+BuildOption(conf):  -DBUILD_TESTING=ON
+BuildOption(conf):  -DMIOPEN_TEST_ALL=ON
+%else
+BuildOption(conf):  -DBUILD_TESTING=OFF
+%endif
+# Disable optional backends not yet packaged on openRuyi
+BuildOption(conf):  -DMIOPEN_USE_COMPOSABLEKERNEL=OFF
+BuildOption(conf):  -DMIOPEN_USE_MLIR=OFF
+
+BuildRequires:  boost-devel
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hipblaslt)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocblas)
+BuildRequires:  cmake(rocrand)
+%if %{with test}
+BuildRequires:  cmake(GTest)
+%endif
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  compiler-rt
+BuildRequires:  half
+BuildRequires:  hipcc
+BuildRequires:  hipblas-common-devel
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(bzip2)
+BuildRequires:  pkgconfig(libzstd)
+BuildRequires:  pkgconfig(nlohmann_json)
+BuildRequires:  pkgconfig(sqlite3)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+# roctracer uses find_path/find_library rather than find_package; no cmake()/pkgconfig() provided
+# FIXME
+BuildRequires:  roctracer-devel
+
+Requires:       cmake(hip)
+Requires:       cmake(rocrand)
+Requires:       gcc-c++
+
+%description
+AMD's library for high performance machine learning primitives.
+MIOpen supports convolution, batch normalization, activation, pooling,
+RNN/LSTM/GRU, and attention/transformer operations for the HIP backend.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep -a
+# clang-tidy is brittle and not needed when rebuilding from a tarball
+sed -i -e 's@clang-tidy@true@' cmake/ClangTidy.cmake
+
+# half_float::detail::expr is not present in all half versions
+sed -i -e 's@std::is_same_v<T, half_float::detail::expr>@0@' test/verify.hpp
+
+# MIOpen tries to download googletest; disable when not needed
+%if %{without test}
+sed -i -e 's@add_subdirectory(test)@#add_subdirectory(test)@' CMakeLists.txt
+sed -i -e 's@add_subdirectory(speedtests)@#add_subdirectory(speedtests)@' CMakeLists.txt
+%endif
+
+# Use the standard data directory for the MIOpen kernel database
+sed -i -e 's@GetLibPath().parent_path() / "share/miopen/db"@"%{_datadir}/miopen/db"@' src/db_path.cpp.in
+
+# -fno-offload-uniform-block is unsupported on this ROCm version
+sed -i -e 's@opts.push_back("-fno-offload-uniform-block");@//opts.push_back("-fno-offload-uniform-block");@' src/comgr.cpp
+
+# Fix the path used to locate the ROCm clang binary at build time
+sed -i -e 's@llvm/bin/clang@bin/clang@' src/hip/hip_build_utils.cpp
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/miopen-hip/LICENSE.md
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/libMIOpen.so.1{,.*}
+%{_libexecdir}/miopen/
+
+%files devel
+%{_datadir}/miopen/
+%{_includedir}/miopen/
+%{_libdir}/cmake/miopen/
+%{_libdir}/libMIOpen.so
+
+%if %{with test}
+%files test
+%{_bindir}/test*
+%endif
+
+%changelog
+%autochangelog
@@ -1,61 +0,0 @@
-From 53d2ea9ad3cc20e1beac2e1c014082c25e221182 Mon Sep 17 00:00:00 2001
-From: Takatoshi Kondo <redboltz@gmail.com>
-Date: Sun, 26 Aug 2018 10:58:47 +0900
-Subject: [PATCH] Fixed #724.
-
-Fixed type mismatch in msgpack_timestamp.
-Added 64bit singed postfix.
---
- include/msgpack/timestamp.h                 | 8 ++++++--
- include/msgpack/v1/adaptor/cpp11/chrono.hpp | 4 ++--
- 2 files changed, 8 insertions(+), 4 deletions(-)
-
-diff --git a/include/msgpack/timestamp.h b/include/msgpack/timestamp.h
-index 4d7df83d..76139312 100644
--- a/include/msgpack/timestamp.h
-+++ b/include/msgpack/timestamp.h
-@@ -28,13 +28,17 @@ static inline bool msgpack_object_to_timestamp(const msgpack_object* obj, msgpac
-     switch (obj->via.ext.size) {
-     case 4:
-         ts->tv_nsec = 0;
-        _msgpack_load32(uint32_t, obj->via.ext.ptr, &ts->tv_sec);
-+        {
-+            uint32_t v;
-+            _msgpack_load32(uint32_t, obj->via.ext.ptr, &v);
-+            ts->tv_sec = v;
-+        }
-         return true;
-     case 8: {
-         uint64_t value;
-         _msgpack_load64(uint64_t, obj->via.ext.ptr, &value);
-         ts->tv_nsec = (uint32_t)(value >> 34);
-        ts->tv_sec = value & 0x00000003ffffffffL;
-+        ts->tv_sec = value & 0x00000003ffffffffLL;
-         return true;
-     }
-     case 12:
-diff --git a/include/msgpack/v1/adaptor/cpp11/chrono.hpp b/include/msgpack/v1/adaptor/cpp11/chrono.hpp
-index 1e08355e..db2035b7 100644
--- a/include/msgpack/v1/adaptor/cpp11/chrono.hpp
-+++ b/include/msgpack/v1/adaptor/cpp11/chrono.hpp
-@@ -41,7 +41,7 @@ struct as<std::chrono::system_clock::time_point> {
-             uint64_t value;
-             _msgpack_load64(uint64_t, o.via.ext.data(), &value);
-             uint32_t nanosec = static_cast<uint32_t>(value >> 34);
-            uint64_t sec = value & 0x00000003ffffffffL;
-+            uint64_t sec = value & 0x00000003ffffffffLL;
-             tp += std::chrono::duration_cast<std::chrono::system_clock::duration>(
-                 std::chrono::nanoseconds(nanosec));
-             tp += std::chrono::seconds(sec);
-@@ -79,7 +79,7 @@ struct convert<std::chrono::system_clock::time_point> {
-             uint64_t value;
-             _msgpack_load64(uint64_t, o.via.ext.data(), &value);
-             uint32_t nanosec = static_cast<uint32_t>(value >> 34);
-            uint64_t sec = value & 0x00000003ffffffffL;
-+            uint64_t sec = value & 0x00000003ffffffffLL;
-             tp += std::chrono::duration_cast<std::chrono::system_clock::duration>(
-                 std::chrono::nanoseconds(nanosec));
-             tp += std::chrono::seconds(sec);
-- 
-2.17.1
-
@@ -1,47 +0,0 @@
-From 232fff18d4f07aa25338da88ce704675f9fea465 Mon Sep 17 00:00:00 2001
-From: Takatoshi Kondo <redboltz@gmail.com>
-Date: Tue, 6 Aug 2024 09:36:04 +0900
-Subject: [PATCH 1/5] Fixed cmake warnings.
-
---
- CMakeLists.txt | 20 +++++++++++++++++---
- 1 file changed, 17 insertions(+), 3 deletions(-)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 8dc6d610a..c75c908f9 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -1,8 +1,7 @@
-CMAKE_MINIMUM_REQUIRED (VERSION 2.8.12)
-+CMAKE_MINIMUM_REQUIRED (VERSION 3.10)
- 
-IF ((CMAKE_VERSION VERSION_GREATER 3.1) OR
-    (CMAKE_VERSION VERSION_EQUAL 3.1))
-    CMAKE_POLICY(SET CMP0054 NEW)
-+IF (MSGPACK_USE_BOOST)
-+    CMAKE_POLICY(SET CMP0167 NEW)
- ENDIF ()
- 
- PROJECT (msgpack)
-@@ -285,7 +285,6 @@
-     # MEMORYCHECK_COMMAND_OPTIONS needs to place prior to CTEST_MEMORYCHECK_COMMAND
-     SET (MEMORYCHECK_COMMAND_OPTIONS "--leak-check=full --show-leak-kinds=definite,possible --error-exitcode=1")
-     FIND_PROGRAM(CTEST_MEMORYCHECK_COMMAND NAMES valgrind)
-    INCLUDE(Dart)
-     ADD_SUBDIRECTORY (test)
- ENDIF ()
- 
-diff --git a/include/msgpack/type.hpp b/include/msgpack/type.hpp
-index 1ab49745f..9ef3e86d3 100644
--- a/include/msgpack/type.hpp
-+++ b/include/msgpack/type.hpp
-@@ -60,7 +60,9 @@
- #if defined(MSGPACK_USE_BOOST)
- 
- #include "adaptor/boost/fusion.hpp"
-+#if !defined(MSGPACK_USE_CPP03)
- #include "adaptor/boost/msgpack_variant.hpp"
-+#endif // !defined(MSGPACK_USE_CPP03)
- #include "adaptor/boost/optional.hpp"
- #include "adaptor/boost/string_ref.hpp"
- #include "adaptor/boost/string_view.hpp"
@@ -1,63 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-Name:           msgpack
-Version:        3.1.0
-Release:        %autorelease
-Summary:        Binary-based efficient object serialization library
-License:        BSL-1.0
-URL:            http://msgpack.org
-#!RemoteAsset
-Source0:        https://github.com/msgpack/msgpack-c/releases/download/cpp-%{version}/%{name}-%{version}.tar.gz
-BuildSystem:    cmake
-
-BuildOption(conf):  -DCMAKE_POLICY_VERSION_MINIMUM=3.5
-
-# https://github.com/msgpack/msgpack-c/commit/53d2ea9ad3cc20e1beac2e1c014082c25e221182
-Patch0:         0001-Fixed-724.patch
-Patch1:         0002-msgpack-cmake4.patch
-
-BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  doxygen
-# for %%check
-BuildRequires:  pkgconfig(gtest)
-BuildRequires:  pkgconfig(zlib)
-
-%description
-MessagePack is a binary-based efficient object serialization
-library. It enables to exchange structured objects between many
-languages like JSON. But unlike JSON, it is very fast and small.
-
-%package        devel
-Summary:        Libraries and header files for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    devel
-Libraries and header files for %{name}
-
-%prep -a
-# gtest 1.17.0 requires at least C++17
-sed -i "s|-std=c++98|-std=gnu++17|g" CMakeLists.txt
-
-%check -p
-# https://github.com/msgpack/msgpack-c/issues/697
-export GTEST_FILTER=-object_with_zone.ext_empty
-
-%files
-%license LICENSE_1_0.txt COPYING
-%doc AUTHORS ChangeLog NOTICE README README.md
-%{_libdir}/*.so.*
-
-%files devel
-%{_includedir}/*
-%{_libdir}/*.so
-%{_libdir}/pkgconfig/msgpack.pc
-%{_libdir}/cmake/msgpack
-
-%changelog
-%autochangelog
@@ -0,0 +1,22 @@
+From b976f614ed2ce3bf98f15e9a93761aafe15ba5a9 Mon Sep 17 00:00:00 2001
+From: Sakura286 <chenxuan@iscas.ac.cn>
+Date: Mon, 16 Mar 2026 15:26:02 +0800
+Subject: [PATCH] disable httpmuxgo121 on newer version of go
+
+---
+ main.go | 2 ++
+ 1 file changed, 2 insertions(+)
+
+diff --git a/main.go b/main.go
+index 650e03a..9a343f3 100644
+--- a/main.go
+++ b/main.go
+@@ -1,3 +1,5 @@
+//go:debug httpmuxgo121=0
+
+ package main
+ 
+ import (
+-- 
+2.53.0
+
@@ -0,0 +1,69 @@
+From eb50883178cae3a721ca8658dde6988ee22c8918 Mon Sep 17 00:00:00 2001
+From: Sakura286 <chenxuan@iscas.ac.cn>
+Date: Mon, 16 Mar 2026 15:45:18 +0800
+Subject: [PATCH] use lib64 instead of lib
+
+---
+ CMakeLists.txt                   | 4 ++--
+ discover/types.go                | 2 +-
+ ml/backend/ggml/ggml/src/ggml.go | 2 +-
+ ml/path.go                       | 2 +-
+ 4 files changed, 5 insertions(+), 5 deletions(-)
+
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 1aa976a..4bd9250 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -37,8 +37,8 @@ if (CMAKE_OSX_ARCHITECTURES MATCHES "x86_64")
+     set(CMAKE_INSTALL_RPATH "@loader_path")
+ endif()
+ 
+-set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib/ollama)
+-set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib/ollama/${OLLAMA_RUNNER_DIR})
+set(OLLAMA_BUILD_DIR ${CMAKE_BINARY_DIR}/lib64/ollama)
+set(OLLAMA_INSTALL_DIR ${CMAKE_INSTALL_PREFIX}/lib64/ollama/${OLLAMA_RUNNER_DIR})
+ 
+ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY         ${OLLAMA_BUILD_DIR})
+ set(CMAKE_RUNTIME_OUTPUT_DIRECTORY_DEBUG   ${OLLAMA_BUILD_DIR})
+diff --git a/discover/types.go b/discover/types.go
+index efc69ec..5848ea4 100644
+--- a/discover/types.go
+++ b/discover/types.go
+@@ -31,7 +31,7 @@ func LogDetails(devices []ml.DeviceInfo) {
+ 	for _, dev := range devices {
+ 		var libs []string
+ 		for _, dir := range dev.LibraryPath {
+-			if strings.Contains(dir, filepath.Join("lib", "ollama")) {
+			if strings.Contains(dir, filepath.Join("lib64", "ollama")) {
+ 				libs = append(libs, filepath.Base(dir))
+ 			}
+ 		}
+diff --git a/ml/backend/ggml/ggml/src/ggml.go b/ml/backend/ggml/ggml/src/ggml.go
+index 7e21591..23f58a1 100644
+--- a/ml/backend/ggml/ggml/src/ggml.go
+++ b/ml/backend/ggml/ggml/src/ggml.go
+@@ -65,7 +65,7 @@ var OnceLoad = sync.OnceFunc(func() {
+ 	case "windows":
+ 		value = filepath.Join(filepath.Dir(exe), "lib", "ollama")
+ 	default:
+-		value = filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
+		value = filepath.Join(filepath.Dir(exe), "..", "lib64", "ollama")
+ 	}
+ 
+ 	// Avoid potentially loading incompatible GGML libraries
+diff --git a/ml/path.go b/ml/path.go
+index ac93af4..3af726c 100644
+--- a/ml/path.go
+++ b/ml/path.go
+@@ -28,7 +28,7 @@ var LibOllamaPath string = func() string {
+ 	case "windows":
+ 		libPath = filepath.Join(filepath.Dir(exe), "lib", "ollama")
+ 	case "linux":
+-		libPath = filepath.Join(filepath.Dir(exe), "..", "lib", "ollama")
+		libPath = filepath.Join(filepath.Dir(exe), "..", "lib64", "ollama")
+ 	case "darwin":
+ 		libPath = filepath.Dir(exe)
+ 	}
+-- 
+2.53.0
+
@@ -0,0 +1,13 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 2820dee..44e0b43 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -28,7 +28,7 @@ set(GGML_CUDA_FA ON)
+ set(GGML_CUDA_COMPRESSION_MODE default)
+ 
+ if((CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_OSX_ARCHITECTURES MATCHES "arm64")
+-    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+"))
+    OR (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_SYSTEM_PROCESSOR MATCHES "arm|aarch64|ARM64|ARMv[0-9]+|riscv64"))
+     set(GGML_CPU_ALL_VARIANTS ON)
+ endif()
+ 
@@ -0,0 +1,25 @@
+From 85aee3e710288343eca393272a418483ac547b83 Mon Sep 17 00:00:00 2001
+From: Sakura286 <chenxuan@iscas.ac.cn>
+Date: Tue, 24 Mar 2026 10:11:07 +0800
+Subject: [PATCH] limit batch size to stablize
+
+---
+ api/types.go | 2 +-
+ 1 file changed, 1 insertion(+), 1 deletion(-)
+
+diff --git a/api/types.go b/api/types.go
+index 3ccf3ce..cba3ffd 100644
+--- a/api/types.go
+++ b/api/types.go
+@@ -896,7 +896,7 @@ func DefaultOptions() Options {
+ 		Runner: Runner{
+ 			// options set when the model is loaded
+ 			NumCtx:    int(envconfig.ContextLength()),
+-			NumBatch:  512,
+			NumBatch:  8,
+ 			NumGPU:    -1, // -1 here indicates that NumGPU should be set dynamically
+ 			NumThread: 0,  // let the runtime decide
+ 			UseMMap:   nil,
+-- 
+2.53.0
+
@@ -0,0 +1,13 @@
+[Unit]
+Description=Ollama Service
+After=network-online.target
+
+[Service]
+ExecStart=/usr/bin/ollama serve
+User=ollama
+Group=ollama
+Restart=always
+RestartSec=3
+
+[Install]
+WantedBy=default.target
@@ -1,6 +1,7 @@
 # SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
 # SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
 # SPDX-FileContributor: misaka00251 <liuxin@iscas.ac.cn>
+# SPDX-FileContributor: Sakura286 <chenxuan@iscas.ac.cn>
 #
 # SPDX-License-Identifier: MulanPSL-2.0

@@ -11,24 +12,24 @@

 # Ollama bundles some ggml libs
 # They should be kept private and the scans of these files should be disabled
-%global __provides_exclude_from ^%{_exec_prefix}/lib/ollama/.*\\.so(\\..*)?$
-%global __requires_exclude ^libggml-base\\.so\\.0\\(\\).*
+%global __provides_exclude lib.*\\.so(\\..*)?
+%global __requires_exclude libggml-.*\\.so(\\..*)?

 Name:           ollama
 Version:        0.13.5
 Release:        %autorelease
 Summary:        Get up and running with OpenAI gpt-oss, DeepSeek-R1, Gemma 3 and other models.
-License:        Apache-2.0 AND MIT
-URL:            https://github.com/ollama/ollama
-#!RemoteAsset
-Source0:        %{url}/archive/refs/tags/v%{version}.tar.gz
+License:        MIT
+URL:            https://ollama.com/
+VCS:            git:https://github.com/ollama/ollama
+#!RemoteAsset:  sha256:6b6bc20a52c11341aa296eecce5ee6782f05815224a4196983b0aa2f1453c19f
+Source0:        https://github.com/ollama/ollama/archive/refs/tags/v%{version}.tar.gz
+Source1:        ollama.service
+Source2:        ollama.sysusers
 BuildSystem:    golang

 BuildOption(prep):  -n %{_name}-%{version}

-Patch0:         0001-ollama-0.14.2_add-riscv.patch
-Patch1:         0002-go-riscv64.patch
-
 BuildRequires:  cmake
 BuildRequires:  fdupes
 BuildRequires:  gcc-c++
@@ -60,12 +61,15 @@ BuildRequires:  go(golang.org/x/tools)
 BuildRequires:  go(gonum.org/v1/gonum)
 BuildRequires:  go(google.golang.org/protobuf)
 BuildRequires:  ninja
-
+BuildRequires:  systemd-rpm-macros
 %if %{with rocm}
 BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(Clang)
 BuildRequires:  cmake(hip)
 BuildRequires:  cmake(hipblas)
 BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(LLD)
+BuildRequires:  cmake(LLVM)
 BuildRequires:  cmake(rocblas)
 BuildRequires:  cmake(rocsolver)
 BuildRequires:  pkgconfig(libdrm_amdgpu)
@@ -73,17 +77,37 @@ BuildRequires:  pkgconfig(libelf)
 BuildRequires:  pkgconfig(numa)
 BuildRequires:  rocm-llvm-macros
 BuildRequires:  rocminfo
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra-devel
 BuildRequires:  compiler-rt
 BuildRequires:  hipcc
-BuildRequires:  lld-devel
-BuildRequires:  llvm-devel
+%endif

+%{?systemd_requires}
+%if %{with rocm}
 Requires:       hipblas
 Requires:       rocblas
 %endif

+%patchlist
+# Ollama vendors ggml code, but it does not sync riscv64 code by default
+# Manually sync riscv64 code here
+0001-ollama-0.14.2_add-riscv.patch
+# Ollama put ggml-cpu code(cpp) inside 'ollama' binary file(go)
+0002-go-riscv64.patch
+# Golang buildsystem on openRuyi use GO11MODULE=off, makes
+# httpmuxgo121=1, which is deprecated in newer version of go
+# Without this patch, ollama cannot provide even the basic http functions
+# https://github.com/jkroepke/openvpn-auth-oauth2/pull/706
+0003-disable-httpmuxgo121-on-newer-version-of-go.patch
+# This patch breaks dlopen of ollama, temporarily disable it
+# Install ollama to /usr/lib as workaround
+# 0004-use-lib64-instead-of-lib.patch
+# GGML_CPU_ALL_VARIANTS only supports x86_64
+0005-disable-cpu-variants.patch
+# Llama.cpp(ggml) on riscv64's ROCm frequently produce nonsense
+# Give parameter '-b 8 -ub 8' can stabilize it
+0006-limit-batch-size-to-stabilize.patch
+
 %description
 Ollama is an open-source platform designed to run large language models locally.
 It allows users to generate text, assist with coding, and create content privately
@@ -95,28 +119,62 @@ rm -rf llama/llama.cpp/vendor

 # Ollama use a mix build of cmake and go.
 # Ollama binary built by go will use dlopen to load *.so built by cmake.
-# Building order is not important.
+# Building order of go/cmake is not important.
 %build -a
-cmake \
-    -B build \
-%if %{with rocm}
-    -DCMAKE_HIP_COMPILER=%rocmllvm_bindir/clang++ \
-    -DAMDGPU_TARGETS=%{rocm_gpu_list_default} \
-%endif
+%cmake \
    -G Ninja \
-    -W no-dev
-cmake --build build --parallel
+    -W no-dev \
+    -DCMAKE_INSTALL_LIBDIR:PATH=lib \
+    -DCMAKE_INSTALL_FULL_LIBDIR:PATH=/usr/lib \
+    -DLIB_INSTALL_DIR:PATH=/usr/lib \
+    -DLIB_SUFFIX= \
+%if %{with rocm}
+    -DCMAKE_HIP_COMPILER=%{rocmllvm_bindir}/clang++ \
+    -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+%endif
+%cmake_build

 %install
 %buildsystem_golang_install
 %cmake_install
+# Remove bundled contents
+rm -rvf %{buildroot}%{_bindir}/lib* \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libamd*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libdrm*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libelf*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libhip*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libhsa*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libnuma* \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libroc*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/libroc*  \
+    %{buildroot}%{_exec_prefix}/lib/ollama/rocblas/
+
+install -p -D -m 0644 %{SOURCE1} %{buildroot}%{_unitdir}/ollama.service
+install -p -D -m 0644 %{SOURCE2} %{buildroot}%{_sysusersdir}/ollama.conf
+# home dir
+mkdir -p %{buildroot}%{_var}/lib/ollama
+
+%pre
+%sysusers_create_package ollama %{SOURCE2}
+
+%preun
+%systemd_preun ollama.service
+
+%post
+%systemd_post ollama.service
+
+%postun
+%systemd_postun_with_restart ollama.service

 %files
 %license LICENSE*
 %doc README*
-%{_bindir}/%{_name}
 %dir %{_exec_prefix}/lib/ollama
+%attr(0755,ollama,ollama) %dir %{_var}/lib/ollama/
+%{_bindir}/ollama
 %{_exec_prefix}/lib/ollama/*
+%{_unitdir}/ollama.service
+%{_sysusersdir}/ollama.conf

 %changelog
 %{?autochangelog}
@@ -0,0 +1 @@
+u ollama - "Runs Ollama" /var/lib/ollama /sbin/nologin
@@ -0,0 +1,56 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global srcname mistral-common
+%global pypi_name mistral_common
+
+Name:           python-%{srcname}
+Version:        1.11.2
+Release:        %autorelease
+Summary:        Library of common utilities for Mistral AI
+License:        Apache-2.0
+URL:            https://github.com/mistralai/mistral-common
+#!RemoteAsset:  sha256:79f68fc2d1190f28637f40e053f919c8c2697e00b2aa679ddee562a95183f4ad
+Source0:        https://files.pythonhosted.org/packages/source/m/%{pypi_name}/%{pypi_name}-%{version}.tar.gz
+# Upstream does not ship the license file in the sdist, fetch it separately
+#!RemoteAsset:  sha256:5ed6f79e77734b5a60740dd821af5ecac9a6f33709c860eea4e20fcb6cca7fcc
+Source1:        https://raw.githubusercontent.com/mistralai/mistral-common/v%{version}/LICENCE
+BuildArch:      noarch
+BuildSystem:    pyproject
+
+BuildOption(install):  %{pypi_name}
+# These modules require the optional "server" extra (fastapi, click, pydantic-settings)
+BuildOption(check):  -e 'mistral_common.experimental.app.*'
+
+BuildRequires:  pyproject-rpm-macros
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  python3dist(pip)
+BuildRequires:  python3dist(setuptools)
+BuildRequires:  python3dist(wheel)
+
+Provides:       python3-%{srcname} = %{version}-%{release}
+%python_provide python3-%{srcname}
+
+%description
+mistral-common is a library of common utilities for Mistral AI, providing
+tokenizers, request and response schemas, and validation helpers used across
+Mistral's models and tooling.
+
+%prep -a
+cp -p %{SOURCE1} LICENCE
+# Relax jsonschema lower bound to match the version available in the repo
+sed -i 's/jsonschema>=4.21.1/jsonschema>=4.17.3/' pyproject.toml
+
+%generate_buildrequires
+%pyproject_buildrequires
+
+%files -f %{pyproject_files}
+%doc README.md
+%license LICENCE
+%{_bindir}/mistral_common
+
+%changelog
+%autochangelog
@@ -0,0 +1,48 @@
+# SPDX-FileCopyrightText: (C) 2025, 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2025, 2026 openRuyi Project Contributors
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global srcname msgpack
+
+Name:           python-%{srcname}
+Version:        1.1.2
+Release:        %autorelease
+Summary:        Python MessagePack (de)serializer
+License:        Apache-2.0
+URL:            https://msgpack.org/
+#!RemoteAsset:  sha256:3b60763c1373dd60f398488069bcdc703cd08a711477b5d480eecc9f9626f47e
+Source0:        https://files.pythonhosted.org/packages/source/m/%{srcname}/%{srcname}-%{version}.tar.gz
+BuildSystem:    pyproject
+
+BuildOption(install):  -l %{srcname}
+
+BuildRequires:  gcc-c++
+BuildRequires:  pyproject-rpm-macros
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  python3dist(pip)
+BuildRequires:  python3dist(setuptools)
+
+Provides:       python3-%{srcname} = %{version}-%{release}
+Provides:       python3-%{srcname}%{?_isa} = %{version}-%{release}
+%python_provide python3-%{srcname}
+
+%description
+MessagePack is a binary-based efficient data interchange format that is
+focused on high performance. It is like JSON, but very fast and small.
+This is a Python (de)serializer for MessagePack.
+
+%prep -a
+# There is a circular dependency with python-msgpack-ext
+rm -rf test/test_timestamp.py
+
+%generate_buildrequires
+%pyproject_buildrequires
+
+%files -f %{pyproject_files}
+%doc README.md
+%license COPYING
+
+%changelog
+%autochangelog
@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global srcname pydantic-extra-types
+%global pypi_name pydantic_extra_types
+
+Name:           python-%{srcname}
+Version:        2.11.1
+Release:        %autorelease
+Summary:        Extra Pydantic types
+License:        MIT
+URL:            https://github.com/pydantic/pydantic-extra-types
+#!RemoteAsset:  sha256:46792d2307383859e923d8fcefa82108b1a141f8a9c0198982b3832ab5ef1049
+Source0:        https://files.pythonhosted.org/packages/source/p/%{pypi_name}/%{pypi_name}-%{version}.tar.gz
+BuildArch:      noarch
+BuildSystem:    pyproject
+
+BuildOption(install):  %{pypi_name}
+# Skip submodules whose optional dependencies are not packaged yet
+BuildOption(check):  -e 'pydantic_extra_types.cron'
+BuildOption(check):  -e 'pydantic_extra_types.mongo_object_id'
+BuildOption(check):  -e 'pydantic_extra_types.pendulum_dt'
+BuildOption(check):  -e 'pydantic_extra_types.phone_numbers'
+BuildOption(check):  -e 'pydantic_extra_types.semantic_version'
+BuildOption(check):  -e 'pydantic_extra_types.semver'
+BuildOption(check):  -e 'pydantic_extra_types.ulid'
+
+BuildRequires:  pyproject-rpm-macros
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  python3dist(hatchling)
+BuildRequires:  python3dist(pip)
+BuildRequires:  python3dist(pycountry)
+BuildRequires:  python3dist(setuptools)
+
+Provides:       python3-%{srcname} = %{version}-%{release}
+%python_provide python3-%{srcname}
+
+%description
+Extra Pydantic types provides a collection of additional field types and
+validators for Pydantic, such as country codes, phone numbers, colors,
+coordinates and currency codes.
+
+%generate_buildrequires
+%pyproject_buildrequires
+
+%files -f %{pyproject_files}
+%doc README.md
+%license LICENSE
+
+%changelog
+%autochangelog
@@ -1,55 +0,0 @@
-From 353550790f659b320ea8753b7d4a6fd701bd1a79 Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:36:36 +0800
-Subject: [PATCH 1/6] fix python shebang
-
---
- Tensile/Configs/miopen/convert_cfg.py | 2 +-
- Tensile/Tests/create_tests.py         | 2 +-
- Tensile/bin/Tensile                   | 2 +-
- Tensile/bin/TensileCreateLibrary      | 2 +-
- 4 files changed, 4 insertions(+), 4 deletions(-)
-
-diff --git a/Tensile/Configs/miopen/convert_cfg.py b/Tensile/Configs/miopen/convert_cfg.py
-index c62d26f..3b5c114 100644
--- a/Tensile/Configs/miopen/convert_cfg.py
-+++ b/Tensile/Configs/miopen/convert_cfg.py
-@@ -1,4 +1,4 @@
-#!/usr/bin/python
-+#!/usr/bin/python3
- 
- ################################################################################
- #
-diff --git a/Tensile/Tests/create_tests.py b/Tensile/Tests/create_tests.py
-index 2b08e3f..94f7345 100755
--- a/Tensile/Tests/create_tests.py
-+++ b/Tensile/Tests/create_tests.py
-@@ -1,4 +1,4 @@
-#!/usr/bin/python
-+#!/usr/bin/python3
- 
- ################################################################################
- #
-diff --git a/Tensile/bin/Tensile b/Tensile/bin/Tensile
-index 1c53682..2ac7d57 100755
--- a/Tensile/bin/Tensile
-+++ b/Tensile/bin/Tensile
-@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
-+#!/usr/bin/python3
- 
- ################################################################################
- #
-diff --git a/Tensile/bin/TensileCreateLibrary b/Tensile/bin/TensileCreateLibrary
-index e90be28..8e966c3 100755
--- a/Tensile/bin/TensileCreateLibrary
-+++ b/Tensile/bin/TensileCreateLibrary
-@@ -1,4 +1,4 @@
-#!/usr/bin/env python3
-+#!/usr/bin/python3
- 
- ################################################################################
- #
-- 
-2.51.0
-
@@ -1,25 +0,0 @@
-From 9d2c031ba924572914b72794f94f1def07aa225c Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:37:40 +0800
-Subject: [PATCH 2/6] fix tensile get path
-
---
- Tensile/cmake/TensileConfig.cmake | 2 +-
- 1 file changed, 1 insertion(+), 1 deletion(-)
-
-diff --git a/Tensile/cmake/TensileConfig.cmake b/Tensile/cmake/TensileConfig.cmake
-index 62682d7..de275b0 100644
--- a/Tensile/cmake/TensileConfig.cmake
-+++ b/Tensile/cmake/TensileConfig.cmake
-@@ -45,7 +45,7 @@ if(NOT DEFINED Tensile_ROOT)
-     if (WIN32)
-         execute_process(COMMAND "${Tensile_PREFIX}/bin/TensileGetPath.exe" OUTPUT_VARIABLE Tensile_ROOT)
-     else()
-        execute_process(COMMAND "${Tensile_PREFIX}/bin/TensileGetPath" OUTPUT_VARIABLE Tensile_ROOT)
-+        execute_process(COMMAND "TensileGetPath" OUTPUT_VARIABLE Tensile_ROOT)
-     endif()
- endif()
- list(APPEND CMAKE_MODULE_PATH "${Tensile_ROOT}/Source/cmake/")
-- 
-2.51.0
-
@@ -1,34 +0,0 @@
-From b1a90000e009daf6c91dc9e0837a36d9f4735a34 Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:39:22 +0800
-Subject: [PATCH 3/6] reduce requirements
-
---
- docs/sphinx/requirements.in  | 1 -
- docs/sphinx/requirements.txt | 2 --
- 2 files changed, 3 deletions(-)
-
-diff --git a/docs/sphinx/requirements.in b/docs/sphinx/requirements.in
-index 4184a90..f818da0 100644
--- a/docs/sphinx/requirements.in
-+++ b/docs/sphinx/requirements.in
-@@ -1,3 +1,2 @@
- rocm-docs-core==1.20.0
- autodoc
-joblib  # Required dependency for API doc-string generation
-diff --git a/docs/sphinx/requirements.txt b/docs/sphinx/requirements.txt
-index e9b7e28..2dd28d9 100644
--- a/docs/sphinx/requirements.txt
-+++ b/docs/sphinx/requirements.txt
-@@ -91,8 +91,6 @@ jinja2==3.1.4
-     # via
-     #   myst-parser
-     #   sphinx
-joblib==1.5.1
-    # via -r requirements.in
- jsonschema==4.23.0
-     # via nbformat
- jsonschema-specifications==2024.10.1
-- 
-2.51.0
-
@@ -1,25 +0,0 @@
-From 5e9b360710400fcace517f055edf54da2ff0e076 Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:41:52 +0800
-Subject: [PATCH 4/6] ignore asm cap cache
-
---
- Tensile/Common.py | 2 +-
- 1 file changed, 1 insertion(+), 1 deletion(-)
-
-diff --git a/Tensile/Common.py b/Tensile/Common.py
-index 86c6c57..376afc4 100644
--- a/Tensile/Common.py
-+++ b/Tensile/Common.py
-@@ -307,7 +307,7 @@ globalParameters["SeparateArchitectures"] = False # write Tensile library metada
- 
- globalParameters["LazyLibraryLoading"] = False # Load library and code object files when needed instead of at startup
- 
-globalParameters["IgnoreAsmCapCache"] = False # Ignore checking for discrepancies between derived and cached asm caps
-+globalParameters["IgnoreAsmCapCache"] = True # Ignore checking for discrepancies between derived and cached asm caps
- 
- globalParameters["ExperimentalLogicDir"] = "/experimental/"
- 
-- 
-2.51.0
-
@@ -1,71 +0,0 @@
-From 746af14c11ee1b455f1adbab8bf6bc3c93fb3fde Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:45:10 +0800
-Subject: [PATCH 5/6] no amdclang when rocm-llvm is unbundled
-
---
- Tensile/Common.py              |  4 ++--
- Tensile/Utilities/Toolchain.py | 10 +++++-----
- 2 files changed, 7 insertions(+), 7 deletions(-)
-
-diff --git a/Tensile/Common.py b/Tensile/Common.py
-index 376afc4..90c579e 100644
--- a/Tensile/Common.py
-+++ b/Tensile/Common.py
-@@ -269,7 +269,7 @@ globalParameters["DictLibraryLogic"] = False
- globalParameters["CurrentISA"] = (0,0,0)
- globalParameters["ROCmAgentEnumeratorPath"] = None      # /opt/rocm/bin/rocm_agent_enumerator
- globalParameters["ROCmSMIPath"] = None                  # /opt/rocm/bin/rocm-smi
-globalParameters["AssemblerPath"] = None                # /opt/rocm/llvm/bin/clang++
-+globalParameters["AssemblerPath"] = "clang++"                # /opt/rocm/llvm/bin/clang++
- globalParameters["WorkingPath"] = os.getcwd()           # path where tensile called from
- globalParameters["IndexChars"] =  "IJKLMNOPQRSTUVWXYZ"  # which characters to use for C[ij]=Sum[k] A[ik]*B[jk]
- globalParameters["ScriptPath"] = os.path.dirname(os.path.realpath(__file__))            # path to Tensile/Tensile.py
-@@ -279,7 +279,7 @@ globalParameters["HipClangVersion"] = "0.0.0"
- globalParameters["RuntimeLanguage"] = "HIP"
- 
- globalParameters["CodeObjectVersion"] = "default"
-globalParameters["CxxCompiler"] = "amdclang++" if os.name != "nt" else "clang++"
-+globalParameters["CxxCompiler"] = "hipcc" if os.name != "nt" else "clang++"
- globalParameters["CCompiler"] = "amdclang" if os.name != "nt" else "clang"
- globalParameters["Architecture"] = "all"
- 
-diff --git a/Tensile/Utilities/Toolchain.py b/Tensile/Utilities/Toolchain.py
-index ee9cbee..e3de82b 100644
--- a/Tensile/Utilities/Toolchain.py
-+++ b/Tensile/Utilities/Toolchain.py
-@@ -106,10 +106,10 @@ def _posixSearchPaths() -> List[Path]:
- 
- 
- class ToolchainDefaults(NamedTuple):
-    CXX_COMPILER = osSelect(linux="amdclang++", windows="clang++.exe")
-    C_COMPILER = osSelect(linux="amdclang", windows="clang.exe")
-+    CXX_COMPILER = osSelect(linux="hipcc", windows="clang++.exe")
-+    C_COMPILER = osSelect(linux="clang", windows="clang.exe")
-     OFFLOAD_BUNDLER = osSelect(linux="clang-offload-bundler", windows="clang-offload-bundler.exe")
-    ASSEMBLER = osSelect(linux="amdclang++", windows="clang++.exe")
-+    ASSEMBLER = osSelect(linux="clang++", windows="clang++.exe")
-     HIP_CONFIG = osSelect(linux="hipconfig", windows="hipconfig")
-     DEVICE_ENUMERATOR = osSelect(linux="rocm_agent_enumerator", windows="hipinfo.exe")
- 
-@@ -132,7 +132,7 @@ def supportedCCompiler(compiler: str) -> bool:
-     Return:
-         If supported True; otherwise, False.
-     """
-    return _supportedComponent(compiler, ["amdclang", "clang", "hipcc"])
-+    return _supportedComponent(compiler, ["clang", "clang", "hipcc"])
- 
- 
- def supportedCxxCompiler(compiler: str) -> bool:
-@@ -144,7 +144,7 @@ def supportedCxxCompiler(compiler: str) -> bool:
-     Return:
-         If supported True; otherwise, False.
-     """
-    return _supportedComponent(compiler, ["amdclang++", "clang++", "hipcc"])
-+    return _supportedComponent(compiler, ["clang++", "clang++", "hipcc"])
- 
- 
- def supportedOffloadBundler(bundler: str) -> bool:
-- 
-2.51.0
-
@@ -1,128 +0,0 @@
-From dfefd5482684998206290e2e62dc0c84dcc7d64e Mon Sep 17 00:00:00 2001
-From: Sakura286 <chenxuan@iscas.ac.cn>
-Date: Fri, 6 Mar 2026 16:49:43 +0800
-Subject: [PATCH 6/6] use system path instead of default
-
---
- Tensile/Common.py                           | 2 +-
- Tensile/Source/CMakeLists.txt               | 4 ++--
- Tensile/Source/FindHIP.cmake                | 4 ++--
- Tensile/Source/cmake/FindROCmSMI.cmake      | 2 +-
- Tensile/Tests/hipModuleLoad_timing/Makefile | 6 +++---
- Tensile/Utilities/Toolchain.py              | 5 ++---
- 6 files changed, 11 insertions(+), 12 deletions(-)
-
-diff --git a/Tensile/Common.py b/Tensile/Common.py
-index 90c579e..3589e01 100644
--- a/Tensile/Common.py
-+++ b/Tensile/Common.py
-@@ -2415,7 +2415,7 @@ def assignGlobalParameters( config, capabilitiesCache: Optional[dict] = None ):
-   if "KeepBuildTmp" in config:
-     globalParameters["KeepBuildTmp"] = config["KeepBuildTmp"]
- 
-  globalParameters["ROCmPath"] = "/opt/rocm"
-+  globalParameters["ROCmPath"] = "/usr"
-   if "ROCM_PATH" in os.environ:
-     globalParameters["ROCmPath"] = os.environ.get("ROCM_PATH")
-   if "TENSILE_ROCM_PATH" in os.environ:
-diff --git a/Tensile/Source/CMakeLists.txt b/Tensile/Source/CMakeLists.txt
-index b96e308..c756756 100644
--- a/Tensile/Source/CMakeLists.txt
-+++ b/Tensile/Source/CMakeLists.txt
-@@ -26,7 +26,7 @@ cmake_minimum_required(VERSION 3.13)
- 
- # Override all paths arguments as they do not work properly
- file(TO_CMAKE_PATH "$ENV{ROCM_PATH}" ROCM_PATH_ENV_VALUE)
-list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH_ENV_VALUE} /opt/rocm)
-+list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH_ENV_VALUE} /usr)
- 
- project(Tensile)
- 
-@@ -65,7 +65,7 @@ CMAKE_DEPENDENT_OPTION(TENSILE_BUILD_CLIENT "Build the benchmarking client" ON
-                         "TENSILE_USE_HIP" OFF)
- 
- if(TENSILE_USE_HIP)
-    find_package(HIP REQUIRED CONFIG PATHS ${ROCM_PATH_ENV_VALUE} /opt/rocm)
-+    find_package(HIP REQUIRED CONFIG PATHS ${ROCM_PATH_ENV_VALUE} /usr)
- endif()
- 
- if(TENSILE_USE_OPENMP)
-diff --git a/Tensile/Source/FindHIP.cmake b/Tensile/Source/FindHIP.cmake
-index d299357..ba8597f 100644
--- a/Tensile/Source/FindHIP.cmake
-+++ b/Tensile/Source/FindHIP.cmake
-@@ -79,7 +79,7 @@ else()
-         hip/hip_runtime.h
-       PATHS
-         ENV HIP_PATH
-        /opt/rocm
-+        /usr
-       PATH_SUFFIXES
-         /include/hip
-         /include
-@@ -98,7 +98,7 @@ else()
-     NAMES  hipcc
-     PATHS
-       ENV HIP_PATH
-          /opt/rocm
-+          /usr
-     PATH_SUFFIXES
-           /bin
-       )
-diff --git a/Tensile/Source/cmake/FindROCmSMI.cmake b/Tensile/Source/cmake/FindROCmSMI.cmake
-index 0498766..071232a 100644
--- a/Tensile/Source/cmake/FindROCmSMI.cmake
-+++ b/Tensile/Source/cmake/FindROCmSMI.cmake
-@@ -24,7 +24,7 @@
- 
- if(NOT ROCM_ROOT)
-     if(NOT ROCM_DIR)
-        set(ROCM_ROOT "/opt/rocm")
-+        set(ROCM_ROOT "/usr")
-     else()
-         set(ROCM_DIR "${ROCM_DIR}/../../..")
-     endif()
-diff --git a/Tensile/Tests/hipModuleLoad_timing/Makefile b/Tensile/Tests/hipModuleLoad_timing/Makefile
-index 671167d..2177143 100644
--- a/Tensile/Tests/hipModuleLoad_timing/Makefile
-+++ b/Tensile/Tests/hipModuleLoad_timing/Makefile
-@@ -22,10 +22,10 @@
- #
- ################################################################################
- 
-CXX?=/opt/rocm/hip/bin/amdclang++
-LIBFLAGS=-L/opt/rocm/hip/lib/
-+CXX?=/usr/bin/amdclang++
-+LIBFLAGS=-L/usr/lib64/
- LIBS=-lamdhip64
-INCFLAGS=-I/opt/rocm/hip/include/
-+INCFLAGS=-I/usr/include/
- 
- hipModuleLoadTiming.out: hipModuleLoadTiming.o
- 	$(CXX) -o $@ $(LIBFLAGS) $^
-diff --git a/Tensile/Utilities/Toolchain.py b/Tensile/Utilities/Toolchain.py
-index e3de82b..e6ee7f3 100644
--- a/Tensile/Utilities/Toolchain.py
-+++ b/Tensile/Utilities/Toolchain.py
-@@ -29,8 +29,8 @@ from subprocess import PIPE, run
- from typing import List, NamedTuple, Union
- from warnings import warn
- 
-DEFAULT_ROCM_BIN_PATH_POSIX = Path("/opt/rocm/bin")
-DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/opt/rocm/lib/llvm/bin")
-+DEFAULT_ROCM_BIN_PATH_POSIX = Path("/usr/bin")
-+DEFAULT_ROCM_LLVM_BIN_PATH_POSIX = Path("/usr/bin")
- DEFAULT_ROCM_BIN_PATH_WINDOWS = Path("C:/Program Files/AMD/ROCm")
- 
- 
-@@ -89,7 +89,6 @@ def _posixSearchPaths() -> List[Path]:
-     if os.environ.get("ROCM_PATH"):
-         for p in os.environ["ROCM_PATH"].split(os.pathsep):
-             searchPaths.append(Path(p) / "bin")
-            searchPaths.append(Path(p) / "lib" / "llvm" / "bin")
- 
-     searchPaths.extend(
-         [
-- 
-2.51.0
-
@@ -1,91 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global srcname tensile
-%global upstreamname Tensile
-%global rocm_version 7.1.1
-
-Name:           python-%{srcname}
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        Tool for creating benchmark-driven backend libraries for GEMMs
-License:        MIT
-URL:            https://github.com/ROCm/Tensile
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    pyproject
-
-BuildOption(install):  -l %{upstreamname}
-
-BuildRequires:  python3-devel
-
-Requires:       cmake-filesystem
-Requires:       hipcc
-Requires:       rocminfo
-Requires:       python3dist(msgpack)
-Requires:       python3dist(pyyaml)
-
-Provides:       python3-%{srcname}
-%python_provide python3-%{srcname}
-
-%patchlist
-0001-fix-python-shebang.patch
-0002-fix-tensile-get-path.patch
-# TODO: joblib is not enabled on openRuyi
-0003-reduce-requirements.patch
-0004-ignore-asm-cap-cache.patch
-# no bundled clang is used on openRuyi
-0005-no-amdclang-when-rocm-llvm-is-unbundled.patch
-# /opt is not used on openRuyi packaging
-0006-use-system-path-instead-of-default.patch
-
-%description
-Tensile is a tool for creating benchmark-driven backend libraries for GEMMs,
-GEMM-like problems (such as batched GEMM), and general N-dimensional tensor
-contractions on a GPU. The Tensile library is mainly used as backend library to
-rocBLAS. Tensile acts as the performance backbone for a wide variety of
-'compute' applications running on AMD GPUs.
-
-%prep -a
-#Fix a few things:
-chmod 755 Tensile/Configs/miopen/convert_cfg.py
-
-%generate_buildrequires
-%pyproject_buildrequires
-
-%install -a
-# /usr/cmake/* -> /usr/lib/cmake/Tensile
-mkdir -p %{buildroot}%{_datadir}/cmake/Tensile
-mv %{buildroot}%{_prefix}/cmake/* %{buildroot}%{_datadir}/cmake/Tensile/
-rm -rf %{buildroot}%{_prefix}/cmake
-
-# Do not distribute broken bins
-rm %{buildroot}%{_bindir}/tensile*
-
-# rm hard links and replace
-rm %{buildroot}%{python3_sitelib}/%{upstreamname}/cmake/*.cmake
-mv %{buildroot}%{_datadir}/cmake/Tensile/*.cmake %{buildroot}%{python3_sitelib}/%{upstreamname}/cmake/
-
-%pyproject_save_files %{upstreamname}
-
-%check
-# 1. tensile requires GPU hardware at runtime
-# 2. optional dependencies (joblib) are intentionally excluded
-
-%files -f %{pyproject_files}
-%doc README.md
-%license LICENSE.md
-# Do not distribute tests
-%exclude %{python3_sitelib}/%{upstreamname}/Tests
-%{_bindir}/Tensile
-%{_bindir}/TensileBenchmarkCluster
-%{_bindir}/TensileCreateLibrary
-%{_bindir}/TensileGetPath
-%{_bindir}/TensileRetuneLibrary
-
-%changelog
-%{?autochangelog}
@@ -0,0 +1,45 @@
+--- a/aten/src/ATen/native/cuda/linalg/BatchLinearAlgebra.cpp
+++ b/aten/src/ATen/native/cuda/linalg/BatchLinearAlgebra.cpp
+@@ -53,11 +53,11 @@
+ } initializer;
+ }  // namespace (anonymous)
+ 
+-#define AT_MAGMA_VERSION MAGMA_VERSION_MAJOR*100 + MAGMA_VERSION_MINOR*10 + MAGMA_VERSION_MICRO
+#define AT_MAGMA_VERSION MAGMA_VERSION_MAJOR*10000 + MAGMA_VERSION_MINOR*100 + MAGMA_VERSION_MICRO
+ 
+-// Check that MAGMA never releases MAGMA_VERSION_MINOR >= 10 or MAGMA_VERSION_MICRO >= 10
+-#if MAGMA_VERSION_MINOR >= 10 || MAGMA_VERSION_MICRO >= 10
+-#error "MAGMA release minor or micro version >= 10, please correct AT_MAGMA_VERSION"
+// Check that MAGMA never releases MAGMA_VERSION_MINOR >= 100 or MAGMA_VERSION_MICRO >= 100
+#if MAGMA_VERSION_MINOR >= 100 || MAGMA_VERSION_MICRO >= 100
+#error "MAGMA release minor or micro version >= 100, please correct AT_MAGMA_VERSION"
+ #endif
+ 
+ #else
+@@ -153,7 +153,7 @@
+     scalar_t** dB_array, magma_int_t lddb, magma_int_t& info,
+     magma_int_t batchsize, const MAGMAQueue& magma_queue, magma_trans_t trans);
+ 
+-#if AT_MAGMA_VERSION >= 254
+#if AT_MAGMA_VERSION >= 20504
+ 
+ template <>
+ void magmaLdlHermitian<double>(
+@@ -209,7 +209,7 @@
+   AT_CUDA_CHECK(cudaGetLastError());
+ }
+ 
+-#endif // AT_MAGMA_VERSION >= 254
+#endif // AT_MAGMA_VERSION >= 20504
+ 
+ template<>
+ void magmaLu<double>(
+@@ -818,7 +818,7 @@
+     // If cusolver and magma 2.5.4+ are both available and hermitian=true,
+     // call magma for complex inputs
+ #ifdef USE_LINALG_SOLVER
+-#if AT_MAGMA_ENABLED() && (AT_MAGMA_VERSION >= 254)
+#if AT_MAGMA_ENABLED() && (AT_MAGMA_VERSION >= 20504)
+       if (LD.is_complex() && hermitian) {
+         return ldl_factor_magma(
+             LD, pivots, info, upper, hermitian);
@@ -0,0 +1,350 @@
+#
+# License Details
+# Main license BSD 3-Clause
+#
+# Apache-2.0
+# android/libs/fbjni/LICENSE
+# android/libs/fbjni/CMakeLists.txt
+# android/libs/fbjni/build.gradle
+# android/libs/fbjni/cxx/fbjni/ByteBuffer.cpp
+# android/libs/fbjni/cxx/fbjni/ByteBuffer.h
+# android/libs/fbjni/cxx/fbjni/Context.h
+# android/libs/fbjni/cxx/fbjni/File.h
+# android/libs/fbjni/cxx/fbjni/JThread.h
+# android/libs/fbjni/cxx/fbjni/NativeRunnable.h
+# android/libs/fbjni/cxx/fbjni/OnLoad.cpp
+# android/libs/fbjni/cxx/fbjni/ReadableByteChannel.cpp
+# android/libs/fbjni/cxx/fbjni/ReadableByteChannel.h
+# android/libs/fbjni/cxx/fbjni/detail/Boxed.h
+# android/libs/fbjni/cxx/fbjni/detail/Common.h
+# android/libs/fbjni/cxx/fbjni/detail/CoreClasses-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/CoreClasses.h
+# android/libs/fbjni/cxx/fbjni/detail/Environment.cpp
+# android/libs/fbjni/cxx/fbjni/detail/Environment.h
+# android/libs/fbjni/cxx/fbjni/detail/Exceptions.cpp
+# android/libs/fbjni/cxx/fbjni/detail/Exceptions.h
+# android/libs/fbjni/cxx/fbjni/detail/FbjniApi.h
+# android/libs/fbjni/cxx/fbjni/detail/Hybrid.cpp
+# android/libs/fbjni/cxx/fbjni/detail/Hybrid.h
+# android/libs/fbjni/cxx/fbjni/detail/Iterator-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/Iterator.h
+# android/libs/fbjni/cxx/fbjni/detail/JWeakReference.h
+# android/libs/fbjni/cxx/fbjni/detail/Log.h
+# android/libs/fbjni/cxx/fbjni/detail/Meta-forward.h
+# android/libs/fbjni/cxx/fbjni/detail/Meta-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/Meta.cpp
+# android/libs/fbjni/cxx/fbjni/detail/Meta.h
+# android/libs/fbjni/cxx/fbjni/detail/MetaConvert.h
+# android/libs/fbjni/cxx/fbjni/detail/ReferenceAllocators-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/ReferenceAllocators.h
+# android/libs/fbjni/cxx/fbjni/detail/References-forward.h
+# android/libs/fbjni/cxx/fbjni/detail/References-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/References.cpp
+# android/libs/fbjni/cxx/fbjni/detail/References.h
+# android/libs/fbjni/cxx/fbjni/detail/Registration-inl.h
+# android/libs/fbjni/cxx/fbjni/detail/Registration.h
+# android/libs/fbjni/cxx/fbjni/detail/SimpleFixedString.h
+# android/libs/fbjni/cxx/fbjni/detail/TypeTraits.h
+# android/libs/fbjni/cxx/fbjni/detail/utf8.cpp
+# android/libs/fbjni/cxx/fbjni/detail/utf8.h
+# android/libs/fbjni/cxx/fbjni/fbjni.cpp
+# android/libs/fbjni/cxx/fbjni/fbjni.h
+# android/libs/fbjni/cxx/lyra/cxa_throw.cpp
+# android/libs/fbjni/cxx/lyra/lyra.cpp
+# android/libs/fbjni/cxx/lyra/lyra.h
+# android/libs/fbjni/cxx/lyra/lyra_breakpad.cpp
+# android/libs/fbjni/cxx/lyra/lyra_exceptions.cpp
+# android/libs/fbjni/cxx/lyra/lyra_exceptions.h
+# android/libs/fbjni/gradle.properties
+# android/libs/fbjni/gradle/android-tasks.gradle
+# android/libs/fbjni/gradle/release.gradle
+# android/libs/fbjni/gradlew
+# android/libs/fbjni/gradlew.bat
+# android/libs/fbjni/host.gradle
+# android/libs/fbjni/java/com/facebook/jni/CppException.java
+# android/libs/fbjni/java/com/facebook/jni/CppSystemErrorException.java
+# android/libs/fbjni/java/com/facebook/jni/DestructorThread.java
+# android/libs/fbjni/java/com/facebook/jni/HybridClassBase.java
+# android/libs/fbjni/java/com/facebook/jni/HybridData.java
+# android/libs/fbjni/java/com/facebook/jni/IteratorHelper.java
+# android/libs/fbjni/java/com/facebook/jni/MapIteratorHelper.java
+# android/libs/fbjni/java/com/facebook/jni/NativeRunnable.java
+# android/libs/fbjni/java/com/facebook/jni/ThreadScopeSupport.java
+# android/libs/fbjni/java/com/facebook/jni/UnknownCppException.java
+# android/libs/fbjni/java/com/facebook/jni/annotations/DoNotStrip.java
+# android/libs/fbjni/scripts/android-setup.sh
+# android/libs/fbjni/scripts/run-host-tests.sh
+# android/libs/fbjni/settings.gradle
+# android/libs/fbjni/test/BaseFBJniTests.java
+# android/libs/fbjni/test/ByteBufferTests.java
+# android/libs/fbjni/test/DocTests.java
+# android/libs/fbjni/test/FBJniTests.java
+# android/libs/fbjni/test/HybridTests.java
+# android/libs/fbjni/test/IteratorTests.java
+# android/libs/fbjni/test/PrimitiveArrayTests.java
+# android/libs/fbjni/test/ReadableByteChannelTests.java
+# android/libs/fbjni/test/jni/CMakeLists.txt
+# android/libs/fbjni/test/jni/byte_buffer_tests.cpp
+# android/libs/fbjni/test/jni/doc_tests.cpp
+# android/libs/fbjni/test/jni/expect.h
+# android/libs/fbjni/test/jni/fbjni_onload.cpp
+# android/libs/fbjni/test/jni/fbjni_tests.cpp
+# android/libs/fbjni/test/jni/hybrid_tests.cpp
+# android/libs/fbjni/test/jni/inter_dso_exception_test_1/Test.cpp
+# android/libs/fbjni/test/jni/inter_dso_exception_test_1/Test.h
+# android/libs/fbjni/test/jni/inter_dso_exception_test_2/Test.cpp
+# android/libs/fbjni/test/jni/inter_dso_exception_test_2/Test.h
+# android/libs/fbjni/test/jni/iterator_tests.cpp
+# android/libs/fbjni/test/jni/modified_utf8_test.cpp
+# android/libs/fbjni/test/jni/no_rtti.cpp
+# android/libs/fbjni/test/jni/no_rtti.h
+# android/libs/fbjni/test/jni/primitive_array_tests.cpp
+# android/libs/fbjni/test/jni/readable_byte_channel_tests.cpp
+# android/libs/fbjni/test/jni/simple_fixed_string_tests.cpp
+# android/libs/fbjni/test/jni/utf16toUTF8_test.cpp
+# android/pytorch_android/host/build.gradle
+# aten/src/ATen/cuda/llvm_basic.cpp
+# aten/src/ATen/cuda/llvm_complex.cpp
+# aten/src/ATen/native/quantized/cpu/qnnpack/confu.yaml
+# aten/src/ATen/native/quantized/cpu/qnnpack/src/requantization/gemmlowp-neon.c
+# aten/src/ATen/native/quantized/cpu/qnnpack/src/requantization/gemmlowp-scalar.h
+# aten/src/ATen/native/quantized/cpu/qnnpack/src/requantization/gemmlowp-sse.h
+# aten/src/ATen/nnapi/codegen.py
+# aten/src/ATen/nnapi/NeuralNetworks.h
+# aten/src/ATen/nnapi/nnapi_wrapper.cpp
+# aten/src/ATen/nnapi/nnapi_wrapper.h
+# binaries/benchmark_args.h
+# binaries/benchmark_helper.cc
+# binaries/benchmark_helper.h
+# binaries/compare_models_torch.cc
+# binaries/convert_and_benchmark.cc
+# binaries/convert_caffe_image_db.cc
+# binaries/convert_db.cc
+# binaries/convert_encoded_to_raw_leveldb.cc
+# binaries/convert_image_to_tensor.cc
+# binaries/core_overhead_benchmark.cc
+# binaries/core_overhead_benchmark_gpu.cc
+# binaries/db_throughput.cc
+# binaries/dump_operator_names.cc
+# binaries/inspect_gpu.cc
+# binaries/load_benchmark_torch.cc
+# binaries/make_cifar_db.cc
+# binaries/make_image_db.cc
+# binaries/make_mnist_db.cc
+# binaries/optimize_for_mobile.cc
+# binaries/parallel_info.cc
+# binaries/predictor_verifier.cc
+# binaries/print_core_object_sizes_gpu.cc
+# binaries/print_registered_core_operators.cc
+# binaries/run_plan.cc
+# binaries/run_plan_mpi.cc
+# binaries/speed_benchmark.cc
+# binaries/speed_benchmark_torch.cc
+# binaries/split_db.cc
+# binaries/tsv_2_proto.cc
+# binaries/tutorial_blob.cc
+# binaries/zmq_feeder.cc
+# c10/test/util/small_vector_test.cpp
+# c10/util/FunctionRef.h
+# c10/util/SmallVector.cpp
+# c10/util/SmallVector.h
+# c10/util/llvmMathExtras.h
+# c10/util/sparse_bitset.h
+# caffe2/contrib/aten/gen_op.py
+# caffe2/contrib/fakelowp/fp16_fc_acc_op.cc
+# caffe2/contrib/fakelowp/fp16_fc_acc_op.h
+# caffe2/contrib/gloo/allgather_ops.cc
+# caffe2/contrib/gloo/allgather_ops.h
+# caffe2/contrib/gloo/reduce_scatter_ops.cc
+# caffe2/contrib/gloo/reduce_scatter_ops.h
+# caffe2/core/hip/common_miopen.h
+# caffe2/core/hip/common_miopen.hip
+# caffe2/core/net_async_tracing.cc
+# caffe2/core/net_async_tracing.h
+# caffe2/core/net_async_tracing_test.cc
+# caffe2/experiments/operators/fully_connected_op_decomposition.cc
+# caffe2/experiments/operators/fully_connected_op_decomposition.h
+# caffe2/experiments/operators/fully_connected_op_decomposition_gpu.cc
+# caffe2/experiments/operators/fully_connected_op_prune.cc
+# caffe2/experiments/operators/fully_connected_op_prune.h
+# caffe2/experiments/operators/fully_connected_op_sparse.cc
+# caffe2/experiments/operators/fully_connected_op_sparse.h
+# caffe2/experiments/operators/funhash_op.cc
+# caffe2/experiments/operators/funhash_op.h
+# caffe2/experiments/operators/sparse_funhash_op.cc
+# caffe2/experiments/operators/sparse_funhash_op.h
+# caffe2/experiments/operators/sparse_matrix_reshape_op.cc
+# caffe2/experiments/operators/sparse_matrix_reshape_op.h
+# caffe2/experiments/operators/tt_contraction_op.cc
+# caffe2/experiments/operators/tt_contraction_op.h
+# caffe2/experiments/operators/tt_contraction_op_gpu.cc
+# caffe2/experiments/operators/tt_pad_op.cc
+# caffe2/experiments/operators/tt_pad_op.h
+# caffe2/experiments/python/SparseTransformer.py
+# caffe2/experiments/python/convnet_benchmarks.py
+# caffe2/experiments/python/device_reduce_sum_bench.py
+# caffe2/experiments/python/funhash_op_test.py
+# caffe2/experiments/python/net_construct_bench.py
+# caffe2/experiments/python/sparse_funhash_op_test.py
+# caffe2/experiments/python/sparse_reshape_op_test.py
+# caffe2/experiments/python/tt_contraction_op_test.py
+# caffe2/experiments/python/tt_pad_op_test.py
+# caffe2/mobile/contrib/libvulkan-stub/include/vulkan/vk_platform.h
+# caffe2/mobile/contrib/libvulkan-stub/include/vulkan/vulkan.h
+# caffe2/mobile/contrib/nnapi/NeuralNetworks.h
+# caffe2/mobile/contrib/nnapi/dlnnapi.c
+# caffe2/mobile/contrib/nnapi/nnapi_benchmark.cc
+# caffe2/observers/profile_observer.cc
+# caffe2/observers/profile_observer.h
+# caffe2/operators/hip/conv_op_miopen.hip
+# caffe2/operators/hip/local_response_normalization_op_miopen.hip
+# caffe2/operators/hip/pool_op_miopen.hip
+# caffe2/operators/hip/spatial_batch_norm_op_miopen.hip
+# caffe2/operators/quantized/int8_utils.h
+# caffe2/operators/stump_func_op.cc
+# caffe2/operators/stump_func_op.cu
+# caffe2/operators/stump_func_op.h
+# caffe2/operators/unique_ops.cc
+# caffe2/operators/unique_ops.cu
+# caffe2/operators/unique_ops.h
+# caffe2/operators/upsample_op.cc
+# caffe2/operators/upsample_op.h
+# caffe2/opt/fusion.h
+# caffe2/python/layers/label_smooth.py
+# caffe2/python/mint/static/css/simple-sidebar.css
+# caffe2/python/modeling/get_entry_from_blobs.py
+# caffe2/python/modeling/get_entry_from_blobs_test.py
+# caffe2/python/modeling/gradient_clipping_test.py
+# caffe2/python/operator_test/unique_ops_test.py
+# caffe2/python/operator_test/upsample_op_test.py
+# caffe2/python/operator_test/weight_scale_test.py
+# caffe2/python/pybind_state_int8.cc
+# caffe2/python/transformations.py
+# caffe2/python/transformations_test.py
+# caffe2/quantization/server/batch_matmul_dnnlowp_op.cc
+# caffe2/quantization/server/batch_matmul_dnnlowp_op.h
+# caffe2/quantization/server/compute_equalization_scale_test.py
+# caffe2/quantization/server/elementwise_linear_dnnlowp_op.cc
+# caffe2/quantization/server/elementwise_linear_dnnlowp_op.h
+# caffe2/quantization/server/elementwise_sum_relu_op.cc
+# caffe2/quantization/server/fb_fc_packed_op.cc
+# caffe2/quantization/server/fb_fc_packed_op.h
+# caffe2/quantization/server/fbgemm_fp16_pack_op.cc
+# caffe2/quantization/server/fbgemm_fp16_pack_op.h
+# caffe2/quantization/server/fully_connected_fake_lowp_op.cc
+# caffe2/quantization/server/fully_connected_fake_lowp_op.h
+# caffe2/quantization/server/int8_gen_quant_params_min_max_test.py
+# caffe2/quantization/server/int8_gen_quant_params_test.py
+# caffe2/quantization/server/int8_quant_scheme_blob_fill_test.py
+# caffe2/quantization/server/spatial_batch_norm_relu_op.cc
+# caffe2/sgd/weight_scale_op.cc
+# caffe2/sgd/weight_scale_op.h
+# caffe2/utils/bench_utils.h
+# functorch/examples/maml_omniglot/maml-omniglot-higher.py
+# functorch/examples/maml_omniglot/maml-omniglot-ptonly.py
+# functorch/examples/maml_omniglot/maml-omniglot-transforms.py
+# functorch/examples/maml_omniglot/support/omniglot_loaders.py
+# modules/detectron/group_spatial_softmax_op.cc
+# modules/detectron/group_spatial_softmax_op.cu
+# modules/detectron/group_spatial_softmax_op.h
+# modules/detectron/ps_roi_pool_op.cc
+# modules/detectron/ps_roi_pool_op.h
+# modules/detectron/roi_pool_f_op.cc
+# modules/detectron/roi_pool_f_op.cu
+# modules/detectron/roi_pool_f_op.h
+# modules/detectron/sample_as_op.cc
+# modules/detectron/sample_as_op.cu
+# modules/detectron/sample_as_op.h
+# modules/detectron/select_smooth_l1_loss_op.cc
+# modules/detectron/select_smooth_l1_loss_op.cu
+# modules/detectron/select_smooth_l1_loss_op.h
+# modules/detectron/sigmoid_cross_entropy_loss_op.cc
+# modules/detectron/sigmoid_cross_entropy_loss_op.cu
+# modules/detectron/sigmoid_cross_entropy_loss_op.h
+# modules/detectron/sigmoid_focal_loss_op.cc
+# modules/detectron/sigmoid_focal_loss_op.cu
+# modules/detectron/sigmoid_focal_loss_op.h
+# modules/detectron/smooth_l1_loss_op.cc
+# modules/detectron/smooth_l1_loss_op.cu
+# modules/detectron/smooth_l1_loss_op.h
+# modules/detectron/softmax_focal_loss_op.cc
+# modules/detectron/softmax_focal_loss_op.cu
+# modules/detectron/softmax_focal_loss_op.h
+# modules/detectron/spatial_narrow_as_op.cc
+# modules/detectron/spatial_narrow_as_op.cu
+# modules/detectron/spatial_narrow_as_op.h
+# modules/detectron/upsample_nearest_op.cc
+# modules/detectron/upsample_nearest_op.h
+# modules/module_test/module_test_dynamic.cc
+# modules/rocksdb/rocksdb.cc
+# scripts/apache_header.txt
+# scripts/apache_python.txt
+# torch/distributions/lkj_cholesky.py
+#
+# Apache 2.0 AND BSD 2-Clause
+# caffe2/operators/deform_conv_op.cu
+#
+# Apache 2.0 AND BSD 2-Clause AND MIT
+# modules/detectron/ps_roi_pool_op.cu
+#
+# Apache 2.0 AND BSD 2-Clause
+# modules/detectron/upsample_nearest_op.cu
+#
+# BSD 0-Clause
+# torch/csrc/utils/pythoncapi_compat.h
+#
+# BSD 2-Clause
+# aten/src/ATen/native/quantized/cpu/qnnpack/deps/clog/LICENSE
+# caffe2/image/transform_gpu.cu
+# caffe2/image/transform_gpu.h
+#
+# BSL-1.0
+# c10/util/flat_hash_map.h
+# c10/util/hash.h
+# c10/util/Optional.h
+# c10/util/order_preserving_flat_hash_map.h
+# c10/util/strong_type.h
+# c10/util/variant.h
+#
+# GPL-3.0-or-later AND MIT
+# c10/util/reverse_iterator.h
+#
+# Khronos
+# These files are for OpenCL, an unused option
+# Replace them later, as-needed with the opencl-headers.rpm
+#
+# caffe2/contrib/opencl/OpenCL/cl.hpp
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl.h
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl.hpp
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl_ext.h
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl_gl.h
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl_gl_ext.h
+# caffe2/mobile/contrib/libopencl-stub/include/CL/cl_platform.h
+# caffe2/mobile/contrib/libopencl-stub/include/CL/opencl.h
+#
+# MIT
+# android/libs/fbjni/googletest-CMakeLists.txt.in
+# c10/util/BFloat16-math.h
+# caffe2/mobile/contrib/libvulkan-stub/include/libvulkan-stub.h
+# caffe2/mobile/contrib/libvulkan-stub/src/libvulkan-stub.c
+# caffe2/onnx/torch_ops/defs.cc
+# cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake
+# cmake/Modules_CUDA_fix/upstream/FindCUDA/parse_cubin.cmake
+# cmake/Modules_CUDA_fix/upstream/FindCUDA/run_nvcc.cmake
+# functorch/einops/_parsing.py
+# test/functorch/test_parsing.py
+# test/functorch/test_rearrange.py
+# third_party/miniz-2.1.0/LICENSE
+# third_party/miniz-2.1.0/miniz.c
+# tools/coverage_plugins_package/setup.py
+# torch/_appdirs.py
+# torch/utils/hipify/hipify_python.py
+#
+# Public Domain
+# caffe2/mobile/contrib/libopencl-stub/LICENSE
+# caffe2/utils/murmur_hash3.cc
+# caffe2/utils/murmur_hash3.h
+#
+# Zlib
+# aten/src/ATen/native/cpu/avx_mathfun.h
+
@@ -0,0 +1,578 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+#
+# Originally extracted from Fedora Project
+# Authors: The Fedora Project Contributors
+
+%global srcname torch
+
+%global toolchain clang
+
+%global pypi_version 2.11.0
+%global miniz_version 3.0.2
+
+# For -test subpackage
+# suitable only for local testing
+# Install and do something like
+#   export LD_LIBRARY_PATH=/usr/lib64/python3.12/site-packages/torch/lib
+#   /usr/lib64/python3.12/site-packages/torch/bin/test_api, test_lazy
+%bcond test 0
+
+%bcond rocm 1
+
+# For testing distributed+rccl etc.
+# TODO: openmpi not included in openRuyi
+%bcond mpi 0
+
+%global _lto_cflags %nil
+
+# Disable dwz with rocm because memory can be exhausted
+%if %{with rocm}
+%define _find_debuginfo_dwz_opts %{nil}
+%endif
+
+# Pytorch third-party buildrequires
+#
+# These system_xxx is kept for debug with some reasons:
+#
+# 1. some package that is not included in openRuyi.
+# 2. some package on openRuyi lack some required component.
+# 3. the corresponding version is mismatched with openRuyi.
+%bcond system_flatbuffers 0
+# Pytorch hardcode httplib to third_party/cpp-httplib
+%bcond system_httplib 0
+# TODO: kineto not included in openruyi
+%bcond system_kineto 0
+# TODO: tensorpipe not included in openRuyi
+%bcond system_tensorpipe 0
+
+Name:           python-%{srcname}
+Version:        %{pypi_version}
+Release:        %autorelease
+Summary:        PyTorch AI/ML framework
+# See license.txt for license details
+License:        BSD-3-Clause AND BSD-2-Clause AND 0BSD AND Apache-2.0 AND MIT AND BSL-1.0 AND GPL-3.0-or-later AND Zlib
+URL:            https://pytorch.org/
+#!RemoteAsset:  sha256:52872a6bbdc42334b00051d88a92f801cfd9be730abdd2b37a2d08996f53bb29
+Source0:        https://github.com/pytorch/pytorch/archive/refs/tags/v%{version}.tar.gz
+%if %{without system_flatbuffers}
+%global flatbuffers_version 24.12.23
+#!RemoteAsset:  sha256:7e2ef35f1af9e2aa0c6a7d0a09298c2cb86caf3d4f58c0658b306256e5bcab10
+Source1:        https://github.com/google/flatbuffers/archive/refs/tags/v%{flatbuffers_version}.tar.gz
+%endif
+%if %{without system_tensorpipe}
+# Developement on tensorpipe has stopped, repo made read only July 1, 2023, this is the last commit
+%global tp_commit 2b4cd91092d335a697416b2a3cb398283246849d
+%global tp_scommit 2b4cd91
+#!RemoteAsset:  sha256:0e85ca56bfe25ed7b3026d2784f716eb10ed1328ade346e3a252814752c57eeb
+Source2:       https://github.com/pytorch/tensorpipe/archive/%{tp_commit}/tensorpipe-%{tp_scommit}.tar.gz
+# The old libuv tensorpipe uses
+#!RemoteAsset:  sha256:6cfeb5f4bab271462b4a2cc77d4ecec847fdbdc26b72019c27ae21509e6f94fa
+Source3:       https://github.com/libuv/libuv/archive/refs/tags/v1.41.0.tar.gz
+# Developement afaik on libnop has stopped, this is the last commit
+%global nop_commit 910b55815be16109f04f4180e9adee14fb4ce281
+%global nop_scommit 910b558
+#!RemoteAsset:  sha256:ec3604671f8ea11aed9588825f9098057ebfef7a8908e97459835150eea9f63a
+Source4:       https://github.com/google/libnop/archive/%{nop_commit}/libnop-%{nop_scommit}.tar.gz
+%endif
+
+%if %{without system_httplib}
+%global hl_commit 4d7c9a788de136071ccf0dd4e96239151e2adadb
+%global hl_scommit 4d7c9a7
+#!RemoteAsset:  sha256:8ecb7bbe844f9b4a1418b8a015d0f815d021d2c0d53291387122cb510c8783ef
+Source5:       https://github.com/yhirose/cpp-httplib/archive/%{hl_commit}/cpp-httplib-%{hl_scommit}.tar.gz
+%endif
+%if %{without system_kineto}
+%global ki_commit 23b5bb5764b3dec988e25c52098407e508d84bb4
+%global ki_scommit 23b5bb5
+#!RemoteAsset:  sha256:5b85352628319e22c48b589d2f423f3761479058f87a3ecc328818f16e4394c6
+Source6:       https://github.com/pytorch/kineto/archive/%{ki_commit}/kineto-%{ki_scommit}.tar.gz
+%endif
+
+%global mslk_commit 3d332d1c0c0ac7765852c97b3979c9ef913e037f
+%global mslk_scommit 3d332d1
+#!RemoteAsset:  sha256:1944e67d1baeffef3bb8f89793ea06e0f05b88aac4d5cd89b4558a21aca6754b
+Source7:       https://github.com/meta-pytorch/MSLK/archive/%{mslk_commit}/MSLK-%{mslk_scommit}.tar.gz
+
+# pytorch upstream issue #173707: libtorch_hip.so references the
+# const_data_ptr / mutable_data_ptr / data_ptr template family with a
+# different (non-SFINAE) mangling than libtorch_cpu.so exports.
+# Appended to aten/src/ATen/core/Tensor.cpp in %prep when rocm is enabled.
+Source8:       pytorch-rocm-symbol-bridge.cpp
+
+# Fix magma version encoding
+# https://github.com/pytorch/pytorch/pull/180388
+Patch0:        0001-pytorch-magma-2.10.0-version-encoding.patch
+
+BuildRequires:  cmake
+BuildRequires:  cmake(concurrentqueue)
+BuildRequires:  cmake(sleef)
+BuildRequires:  cpuinfo
+# Although eigen3 enabled on openruyi, it cannot be detected during conf
+# TODO: Fix this
+BuildRequires:  eigen3
+BuildRequires:  foxi-devel
+BuildRequires:  libomp-devel
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(fmt)
+BuildRequires:  pkgconfig(nlohmann_json)
+BuildRequires:  pkgconfig(numa)
+BuildRequires:  pkgconfig(openblas64)
+BuildRequires:  pkgconfig(protobuf)
+BuildRequires:  pkgconfig(valgrind)
+BuildRequires:  pocketfft-devel
+BuildRequires:  pthreadpool-devel
+BuildRequires:  fp16-devel
+BuildRequires:  fxdiv-devel
+BuildRequires:  psimd-devel
+BuildRequires:  xnnpack-devel = 0+git20260211.312eb7e
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  python3dist(filelock)
+BuildRequires:  python3dist(jinja2)
+BuildRequires:  python3dist(networkx)
+BuildRequires:  python3dist(numpy)
+BuildRequires:  python3dist(pip)
+BuildRequires:  python3dist(pybind11)
+BuildRequires:  python3dist(pyyaml)
+BuildRequires:  python3dist(setuptools)
+BuildRequires:  python3dist(sympy)
+BuildRequires:  python3dist(typing-extensions)
+
+%if %{with system_httplib}
+BuildRequires:  cmake(httplib)
+%endif
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  libstdc++-devel
+BuildRequires:  compiler-rt
+BuildRequires:  cmake(LLVM)
+BuildRequires:  lld
+
+BuildRequires:  cmake(ONNX)
+BuildRequires:  cmake(onnxruntime)
+
+%if %{with mpi}
+BuildRequires:  openmpi-devel
+%endif
+
+%if %{with system_flatbuffers}
+BuildRequires:  pkgconfig(flatbuffers)
+%endif
+
+%if %{with rocm}
+BuildRequires:  cmake(hipblas)
+BuildRequires:  cmake(hipblaslt)
+BuildRequires:  cmake(hipcub)
+BuildRequires:  cmake(hipfft)
+BuildRequires:  cmake(hiprand)
+BuildRequires:  cmake(hipsparse)
+BuildRequires:  cmake(hipsparselt)
+BuildRequires:  cmake(hipsolver)
+BuildRequires:  cmake(miopen)
+BuildRequires:  cmake(rocblas)
+BuildRequires:  cmake(rocrand)
+BuildRequires:  cmake(rocfft)
+BuildRequires:  cmake(rccl)
+BuildRequires:  cmake(rocprim)
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(rocm-core)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocsolver)
+BuildRequires:  cmake(rocm_smi)
+BuildRequires:  cmake(rocthrust)
+BuildRequires:  pkgconfig(magma)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-llvm-macros
+BuildRequires:  roctracer-devel
+%endif
+
+Requires:       python3dist(dill)
+Requires:       python3dist(pyyaml)
+%if %{with rocm}
+Requires:       amdsmi
+%endif
+
+# As convention
+Provides:       pytorch = %{version}-%{release}
+Provides:       python3-%{srcname} = %{version}-%{release}
+Provides:       python3-%{srcname}%{?_isa} = %{version}-%{release}
+%python_provide python3-%{srcname}
+
+%description
+PyTorch is a Python package that provides two high-level features:
+
+ * Tensor computation (like NumPy) with strong GPU acceleration
+ * Deep neural networks built on a tape-based autograd system
+
+You can reuse your favorite Python packages such as NumPy, SciPy,
+and Cython to extend PyTorch when needed.
+
+%prep
+%autosetup -p1 -n pytorch-%{version}
+
+# GitHub release tarballs identify the version as an alpha, so replace that
+echo "%{pypi_version}" > version.txt
+
+# Remove bundled egg-info
+rm -rf %{srcname}.egg-info
+
+%if %{without system_flatbuffers}
+tar xf %{SOURCE1}
+rm -rf third_party/flatbuffers/*
+cp -r flatbuffers-%{flatbuffers_version}/* third_party/flatbuffers/
+%endif
+
+%if %{without system_tensorpipe}
+tar xf %{SOURCE2}
+rm -rf third_party/tensorpipe/*
+cp -r tensorpipe-*/* third_party/tensorpipe/
+tar xf %{SOURCE3}
+rm -rf third_party/tensorpipe/third_party/libuv/*
+cp -r libuv-*/* third_party/tensorpipe/third_party/libuv/
+tar xf %{SOURCE4}
+rm -rf third_party/tensorpipe/third_party/libnop/*
+cp -r libnop-*/* third_party/tensorpipe/third_party/libnop/
+
+# gcc 15 include cstdint
+sed -i '/#include <tensorpipe.*/a#include <cstdint>' third_party/tensorpipe/tensorpipe/common/allocator.h
+sed -i '/#include <tensorpipe.*/a#include <cstdint>' third_party/tensorpipe/tensorpipe/common/memory.h
+%endif
+
+%if %{without system_httplib}
+tar xf %{SOURCE5}
+rm -rf third_party/cpp-httplib/*
+cp -r cpp-httplib-*/* third_party/cpp-httplib/
+%endif
+
+%if %{without system_kineto}
+tar xf %{SOURCE6}
+rm -rf third_party/kineto/*
+cp -r kineto-*/* third_party/kineto/
+%endif
+
+tar xf %{SOURCE7}
+rm -rf third_party/mslk/*
+cp -r MSLK-*/* third_party/mslk/
+
+# Adjust for amd gpu targets currently supported
+# only gfx1100 supported on openruyi
+sed -i -e 's@"gfx90a", "gfx942",@@' aten/src/ATen/native/cuda/Blas.cpp
+sed -i -e 's@"gfx1100", "gfx1101", "gfx1200", "gfx1201", "gfx908"@"gfx1100", "gfx1101",@' aten/src/ATen/native/cuda/Blas.cpp
+sed -i -e 's@"gfx950", "gfx1150", "gfx1151"@@' aten/src/ATen/native/cuda/Blas.cpp
+
+# Need to pip this
+sed -i -e '/fsspec/d' setup.py
+
+# Use system sympy
+sed -i -e 's@sympy==1.13.1@sympy>=1.13.1@' setup.py
+
+# A new dependency
+# Connected to USE_FLASH_ATTENTION, since this is off, do not need it
+sed -i -e '/aotriton.cmake/d' cmake/Dependencies.cmake
+# Compress hip
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc --offload-compress@' cmake/Dependencies.cmake
+# Silence noisy warning
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -Wno-pass-failed@' cmake/Dependencies.cmake
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -Wno-unused-command-line-argument@' cmake/Dependencies.cmake
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -Wno-unused-result@' cmake/Dependencies.cmake
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -Wno-deprecated-declarations@' cmake/Dependencies.cmake
+# Fix: error: branch size exceeds simm16 (AMDGPUAsmBackend.cpp)
+# -amdgpu-s-branch-bits=15(default is 16) and -amdgpu-long-branch-factor=2 are needed to avoid 'branch size exceed simm16' error
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -mllvm --amdgpu-s-branch-bits=15@' cmake/Dependencies.cmake
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc -mllvm --amdgpu-long-branch-factor=2@' cmake/Dependencies.cmake
+
+# Use parallel jobs for GPU offload compilation
+sed -i -e 's@HIP_CLANG_FLAGS -fno-gpu-rdc@HIP_CLANG_FLAGS -fno-gpu-rdc --offload-jobs=8@' cmake/Dependencies.cmake
+# Need to link with librocm_smi64 (intra_node_comm.cpp calls rsmi_init /
+# rsmi_is_P2P_accessible). The target string is "hiprtc::hiprtc" — the previous
+# pattern "hipzrtc::hiprtc" had a stray 'z' so the sed was a no-op and
+# libtorch_hip.so ended up with an undefined rsmi_init symbol.
+sed -i -e 's@hiprtc::hiprtc@hiprtc::hiprtc rocm_smi64@' cmake/Dependencies.cmake
+
+# No third_party fmt, use system
+sed -i -e 's@fmt::fmt-header-only@fmt@' CMakeLists.txt
+sed -i -e 's@fmt::fmt-header-only@fmt@' aten/src/ATen/CMakeLists.txt
+sed -i -e 's@list(APPEND ATen_HIP_INCLUDE $<TARGET_PROPERTY:fmt,INTERFACE_INCLUDE_DIRECTORIES>)@@' aten/src/ATen/CMakeLists.txt
+
+sed -i -e 's@fmt::fmt-header-only@fmt@' third_party/kineto/libkineto/CMakeLists.txt
+sed -i -e 's@fmt::fmt-header-only@fmt@' c10/CMakeLists.txt
+sed -i -e 's@fmt::fmt-header-only@fmt@' torch/CMakeLists.txt
+sed -i -e 's@fmt::fmt-header-only@fmt@' cmake/Dependencies.cmake
+sed -i -e 's@fmt::fmt-header-only@fmt@' caffe2/CMakeLists.txt
+
+sed -i -e 's@add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)@#add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)@' cmake/Dependencies.cmake
+sed -i -e 's@set_target_properties(fmt-header-only PROPERTIES INTERFACE_COMPILE_FEATURES "")@#set_target_properties(fmt-header-only PROPERTIES INTERFACE_COMPILE_FEATURES "")@' cmake/Dependencies.cmake
+sed -i -e 's@list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)@#list(APPEND Caffe2_DEPENDENCY_LIBS fmt::fmt-header-only)@' cmake/Dependencies.cmake
+
+# No third_party FXdiv
+sed -i -e 's@if(NOT TARGET fxdiv)@if(MSVC AND USE_XNNPACK)@' caffe2/CMakeLists.txt
+sed -i -e 's@TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)@#TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)@' caffe2/CMakeLists.txt
+
+# https://github.com/pytorch/pytorch/issues/149803
+# Tries to checkout nccl
+sed -i -e 's@    checkout_nccl()@    True@' tools/build_pytorch_libs.py
+
+# Disable the use of check_submodule's in the setup.py, we are a tarball, not a git repo
+sed -i -e 's@check_submodules()$@#check_submodules()@' setup.py
+
+# Release comes fully loaded with third party src
+# Remove what we can
+#
+# For 2.1 this is all but miniz-2.1.0
+# Instead of building as a library, caffe2 reaches into
+# the third_party dir to compile the file.
+# mimiz is licensed MIT
+# https://github.com/richgel999/miniz/blob/master/LICENSE
+mv third_party/miniz-%{miniz_version} .
+#
+# setup.py depends on this script
+mv third_party/build_bundled.py .
+
+%if %{without system_flatbuffers}
+# Need the just untarred flatbuffers/flatbuffers.h
+mv third_party/flatbuffers .
+%endif
+
+%if %{without system_tensorpipe}
+mv third_party/tensorpipe .
+%endif
+
+%if %{without system_httplib}
+mv third_party/cpp-httplib .
+%endif
+
+%if %{without system_kineto}
+mv third_party/kineto .
+%endif
+
+mv third_party/mslk .
+
+# Remove everything
+rm -rf third_party/*
+# Put stuff back
+mv build_bundled.py third_party
+mv miniz-%{miniz_version} third_party
+
+%if %{without system_flatbuffers}
+mv flatbuffers third_party
+%endif
+
+%if %{without system_tensorpipe}
+mv tensorpipe third_party
+%endif
+
+%if %{without system_httplib}
+mv cpp-httplib third_party
+%endif
+
+%if %{without system_kineto}
+mv kineto third_party
+%endif
+
+mv mslk third_party
+
+# Fake out pocketfft, and system header will be used
+mkdir third_party/pocketfft
+cp /usr/include/pocketfft_hdronly.h third_party/pocketfft/
+
+# Use the system valgrind headers
+mkdir third_party/valgrind-headers
+cp %{_includedir}/valgrind/* third_party/valgrind-headers
+
+# Fix installing to /usr/lib64
+sed -i -e 's@DESTINATION ${PYTHON_LIB_REL_PATH}@DESTINATION ${CMAKE_INSTALL_PREFIX}/${PYTHON_LIB_REL_PATH}@' caffe2/CMakeLists.txt
+
+# reenable foxi linking
+sed -i -e 's@list(APPEND Caffe2_DEPENDENCY_LIBS foxi_loader)@#list(APPEND Caffe2_DEPENDENCY_LIBS foxi_loader)@' cmake/Dependencies.cmake
+
+%if %{without system_tensorpipe}
+# cmake version changed
+sed -i -e 's@cmake_minimum_required(VERSION 3.4)@cmake_minimum_required(VERSION 3.5)@' third_party/tensorpipe/third_party/libuv/CMakeLists.txt
+sed -i -e 's@cmake_minimum_required(VERSION 3.4)@cmake_minimum_required(VERSION 3.5)@' libuv*/CMakeLists.txt
+%endif
+
+%if %{with rocm}
+# Fix: hipOccupancyMaxActiveBlocksPerMultiprocessor is overloaded in new ROCm,
+# force using hipModuleOccupancyMaxActiveBlocksPerMultiprocessor
+sed -i -e 's/TORCH_HIP_VERSION < 305/TORCH_HIP_VERSION < 305 \&\& TORCH_HIP_VERSION > 0/' \
+    aten/src/ATen/cuda/nvrtc_stub/ATenNVRTC.h
+# pytorch upstream issue #173707 (gemm/bgemm variant):
+# clang 21 mangles the instantiation-dependent SFINAE non-type template parameter
+#   typename std::enable_if<...,Dtype>::type* = nullptr
+# of at::cuda::blas::gemm/bgemm differently at an explicit specialization (the
+# definition, Tn...enable_if form) than at a deduced call site (the reference,
+# ...IffLPf0E... form), so libtorch_hip.so fails to dlopen with e.g.
+#   undefined symbol: _ZN2at4cuda4blas4gemmIffLPf0EEEvcclllNS_10OpMathTypeIT_E4typeEPKS5_lS9_lS7_PT0_l
+# Every real dtype is provided by an explicit specialization, so the SFINAE guard
+# is redundant: drop it so the two overloads collapse to one primary template and
+# clang emits a single consistent mangling everywhere. Must run before hipify.
+sed -i \
+    -e 's/, typename std::enable_if<!CUDABLAS_GEMM_DTYPE_IS_FLOAT_TYPE_AND_C_DTYPE_IS_FLOAT, Dtype>::type\* = nullptr>/>/g' \
+    -e 's/, typename std::enable_if<CUDABLAS_GEMM_DTYPE_IS_FLOAT_TYPE_AND_C_DTYPE_IS_FLOAT, Dtype>::type\* = nullptr>/>/g' \
+    aten/src/ATen/cuda/CUDABlas.h
+# hipify
+./tools/amd_build/build_amd.py
+# use any hip, correct CMAKE_MODULE_PATH
+sed -i -e 's@lib/cmake/hip@lib64/cmake/hip@' cmake/public/LoadHIP.cmake
+sed -i -e 's@HIP 1.0@HIP MODULE@'            cmake/public/LoadHIP.cmake
+# silence an assert
+# sed -i -e '/qvalue = std::clamp(qvalue, qmin, qmax);/d' aten/src/ATen/native/cuda/IndexKernel.cu
+
+# Append ROCm symbol bridge — see Source8 header for full context.
+# Without this, libtorch_hip.so dlopen fails on:
+#   undefined symbol: _ZNK2at10TensorBase14const_data_ptrI*Li0EEEPK*v
+cat %{SOURCE8} >> aten/src/ATen/core/Tensor.cpp
+%endif
+
+# moodycamel include path needs adjusting to use the system's
+sed -i -e 's@${PROJECT_SOURCE_DIR}/third_party/concurrentqueue@/usr/include/concurrentqueue@' cmake/Dependencies.cmake
+
+%build
+# Control the number of jobs
+# The build can fail if too many threads exceed the physical memory
+# Run at least one thread, more if CPU & memory resources are available.
+COMPILE_JOBS=`nproc`
+if [ ${COMPILE_JOBS}x = x ]; then
+    COMPILE_JOBS=1
+fi
+# Take into account memory usage per core, do not thrash real memory
+# TraceType/VariableType files can consume 4GB+ per compilation unit
+# Use a more conservative estimate: 4GB per job for safety
+BUILD_MEM=4
+MEM_KB=0
+MEM_KB=`cat /proc/meminfo | grep MemTotal | awk '{ print $2 }'`
+MEM_MB=`eval "expr ${MEM_KB} / 1024"`
+MEM_GB=`eval "expr ${MEM_MB} / 1024"`
+COMPILE_JOBS_MEM=`eval "expr 1 + ${MEM_GB} / ${BUILD_MEM}"`
+if [ "$COMPILE_JOBS_MEM" -lt "$COMPILE_JOBS" ]; then
+    COMPILE_JOBS=$COMPILE_JOBS_MEM
+fi
+# Ensure at least 2 jobs to avoid single-threading the large files
+if [ "$COMPILE_JOBS" -lt 2 ]; then
+    COMPILE_JOBS=2
+fi
+export MAX_JOBS=$COMPILE_JOBS
+
+# For verbose cmake output
+# export VERBOSE=ON
+# For verbose linking
+# export CMAKE_SHARED_LINKER_FLAGS=-Wl,--verbose
+
+# Manually set this hardening flag
+export CMAKE_EXE_LINKER_FLAGS=-pie
+export BUILD_CUSTOM_PROTOBUF=OFF
+export BUILD_NVFUSER=OFF
+export BUILD_SHARED_LIBS=ON
+export BUILD_TEST=OFF
+# Use Release instead of RelWithDebInfo to reduce compile time and memory
+# for huge generated files like TraceType/VariableType (saves ~30% compile time)
+export CMAKE_BUILD_TYPE=Release
+export CMAKE_FIND_PACKAGE_PREFER_CONFIG=ON
+export CAFFE2_LINK_LOCAL_PROTOBUF=OFF
+export INTERN_BUILD_MOBILE=OFF
+export USE_DISTRIBUTED=OFF
+export USE_CUDA=OFF
+export USE_FAKELOWP=OFF
+export USE_FBGEMM=OFF
+export USE_FLASH_ATTENTION=OFF
+export USE_GLOO=OFF
+export USE_ITT=OFF
+export USE_KINETO=OFF
+export USE_KLEIDIAI=OFF
+export USE_LITE_INTERPRETER_PROFILER=OFF
+export USE_LITE_PROTO=OFF
+export USE_MAGMA=OFF
+export USE_MEM_EFF_ATTENTION=OFF
+export USE_MKLDNN=OFF
+export USE_MPI=OFF
+export USE_MSLK=OFF
+export USE_NCCL=OFF
+export USE_NNPACK=OFF
+export USE_NUMPY=ON
+export USE_OPENMP=ON
+export USE_PYTORCH_QNNPACK=OFF
+export USE_ROCM=OFF
+export USE_SYSTEM_SLEEF=ON
+export USE_SYSTEM_EIGEN_INSTALL=ON
+export USE_SYSTEM_ONNX=ON
+export USE_SYSTEM_PYBIND11=ON
+export USE_SYSTEM_LIBS=OFF
+export USE_SYSTEM_NCCL=OFF
+export USE_XNNPACK=OFF
+export USE_XPU=OFF
+export USE_SYSTEM_PTHREADPOOL=ON
+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_DISTRIBUTED=ON
+export USE_TENSORPIPE=ON
+%if %{without system_tensorpipe}
+export TP_BUILD_LIBUV=OFF
+%endif
+
+%if %{with mpi}
+export USE_MPI=ON
+%endif
+
+%if %{with rocm}
+export USE_ROCM=ON
+export USE_ROCM_CK_SDPA=OFF
+export USE_ROCM_CK_GEMM=OFF
+export USE_FBGEMM_GENAI=OFF
+
+export USE_MAGMA=ON
+export HIP_PATH=`hipconfig -p`
+export ROCM_PATH=`hipconfig -R`
+
+# pytorch uses clang, not hipcc
+export HIP_CLANG_PATH=%{rocmllvm_bindir}
+export PYTORCH_ROCM_ARCH=%{rocm_gpu_list_default}
+
+export CMAKE_NO_SYSTEM_FROM_IMPORTED=ON
+
+# export CMAKE_BUILD_TYPE=Debug
+%endif
+
+export CMAKE_CXX_IMPLICIT_INCLUDE_DIRECTORIES="/usr/include"
+export CMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES="/usr/include"
+
+export LDFLAGS="-fuse-ld=lld %{?__global_ldflags}"
+export CMAKE_LIBRARY_PATH=/usr/lib64
+export CMAKE_PREFIX_PATH="/usr:/usr/lib64/cmake:/usr/lib/python3.13/site-packages"
+
+%pyproject_wheel
+
+%install
+%if %{with rocm}
+export USE_ROCM=ON
+export USE_ROCM_CK=OFF
+export HIP_PATH=`hipconfig -p`
+export ROCM_PATH=`hipconfig -R`
+
+# pytorch uses clang, not hipcc
+export HIP_CLANG_PATH=%{rocmllvm_bindir}
+export PYTORCH_ROCM_ARCH=%{rocm_gpu_list_default}
+%endif
+
+%pyproject_install
+%pyproject_save_files '*torch*'
+
+%check
+# Not working yet
+
+%files
+%license LICENSE
+%doc README.md
+%{_bindir}/torchrun
+%{python3_sitearch}/%{srcname}*
+%{python3_sitearch}/functorch
+
+%changelog
+%{?autochangelog}
@@ -0,0 +1,183 @@
+// === openRuyi ROCm symbol bridge for pytorch 2.11 / clang 21 ===
+//
+// Appended to aten/src/ATen/core/Tensor.cpp by python-torch.spec.
+//
+// Background:
+//   libtorch_hip.so references the TensorBase data-pointer template family
+//   (const_data_ptr / mutable_data_ptr / data_ptr) using a NON-SFINAE
+//   mangling form (...Li0EEE... — only the non-type template parameter
+//   value is encoded). The explicit specialisations emitted from
+//   TensorMethods.cpp into libtorch_cpu.so may or may not be mangled
+//   the same way depending on clang's handling of SFINAE NTTPs in this
+//   specific clang/ROCm/arch combination (we have observed both: cpu
+//   exporting the Li0E form, and cpu only exporting the Tn...enable_if
+//   form). No link-time error in the latter case thanks to lld's
+//   --allow-shlib-undefined → runtime dlopen reports
+//   "undefined symbol: _ZNK2at10TensorBase14const_data_ptrI*Li0EEEPK*v"
+//   when libtorch_hip.so is loaded.
+//
+// Reference: https://github.com/pytorch/pytorch/issues/173707
+//   (closed as not planned; pytorch treats this as a clang/ROCm gap)
+//
+// Bridge strategy:
+//   Provide every plausibly-missing mangling as a weak free function
+//   linked into libtorch_cpu.so with default visibility. Where the cpp
+//   specialisation already provides the same mangled name as a strong
+//   symbol, the linker discards the bridge weak symbol and uses the
+//   cpp version (which preserves the runtime check_type call). Where
+//   the cpp does NOT emit the same mangling, the bridge weak symbol
+//   fills the gap so libtorch_hip.so dlopen resolves. Each bridge body
+//   delegates to the non-templated public accessor on TensorBase,
+//   which returns the raw underlying data pointer.
+//
+// Semantics note:
+//   The non-templated TensorBase::const_data_ptr() / mutable_data_ptr() /
+//   data_ptr() skip the scalar-type runtime check that the templated
+//   specialisations perform. In practice this only matters for HIP code
+//   paths that already dispatched on dtype before reaching this call —
+//   which is the common case in ATen kernels.
+//
+// gemm stubs from the upstream issue are intentionally NOT included:
+//   `nm -DC --undefined-only libtorch_hip.so` on this build shows no
+//   undefined at::cuda::blas::gemm symbols, and the upstream stubs have
+//   empty bodies (silent functional failure if ever called).
+
+#include <ATen/core/TensorBase.h>
+#include <c10/util/BFloat16.h>
+#include <c10/util/Float8_e4m3fn.h>
+#include <c10/util/Float8_e4m3fnuz.h>
+#include <c10/util/Float8_e5m2.h>
+#include <c10/util/Float8_e5m2fnuz.h>
+#include <c10/util/Float8_e8m0fnu.h>
+#include <c10/util/Half.h>
+#include <c10/util/complex.h>
+#include <c10/util/qint8.h>
+#include <c10/util/qint32.h>
+#include <c10/util/quint8.h>
+
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wmissing-prototypes"
+#pragma GCC visibility push(default)
+
+extern "C" {
+
+// `weak` is essential: the cpp's TensorMethods.cpp specialisation emits
+// the same mangled name as a strong global on some build configurations.
+// Without `weak` the link step fails with a duplicate-symbol error.
+#define BRIDGE_READ(MangledName) \
+  __attribute__((weak, visibility("default"))) \
+  const void* MangledName(const at::TensorBase* t) { return t->const_data_ptr(); }
+
+#define BRIDGE_WRITE(MangledName) \
+  __attribute__((weak, visibility("default"))) \
+  void* MangledName(const at::TensorBase* t) { return t->mutable_data_ptr(); }
+
+// ---- const_data_ptr<T, 0>  (non-const T, plain Li0E form) ----
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIaLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIbLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIdLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIfLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIhLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIiLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIjLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIlLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrImLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIsLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrItLi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c104HalfELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c108BFloat16ELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c107complexIdEELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c107complexIfEELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c107complexINS2_4HalfEEELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c1011Float8_e5m2ELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c1013Float8_e4m3fnELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c1014Float8_e8m0fnuELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c1015Float8_e4m3fnuzELi0EEEPKT_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c1015Float8_e5m2fnuzELi0EEEPKT_v)
+
+// ---- const_data_ptr<KT, 0>  (const-qualified T, plain Li0E form) ----
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKaLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKbLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKdLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKfLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKhLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKiLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKjLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKlLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKmLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKsLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKtLi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c104HalfELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c108BFloat16ELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c107complexIdEELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c107complexIfEELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c107complexINS2_4HalfEEELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c105qint8ELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c106qint32ELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c106quint8ELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c1011Float8_e5m2ELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c1013Float8_e4m3fnELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c1014Float8_e8m0fnuELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c1015Float8_e4m3fnuzELi0EEEPKNSt12remove_constIT_E4typeEv)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIKN3c1015Float8_e5m2fnuzELi0EEEPKNSt12remove_constIT_E4typeEv)
+
+// ---- const_data_ptr<T, Tn enable_if<!is_const_v<T>>... 0>  (SFINAE form, non-const T) ----
+// libtorch_hip.so emits these for a few primitives even when most TUs use the Li0E form.
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIdTnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS3_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIfTnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS3_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIiTnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS3_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIlTnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS3_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c104HalfETnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS5_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c108BFloat16ETnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS5_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c107complexIdEETnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS6_v)
+BRIDGE_READ(_ZNK2at10TensorBase14const_data_ptrIN3c107complexIfEETnNSt9enable_ifIXntsr3stdE10is_const_vIT_EEiE4typeELi0EEEPKS6_v)
+
+// ---- mutable_data_ptr<T> ----
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIaEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIbEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIdEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIfEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIhEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIiEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIjEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIlEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrImEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIsEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrItEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c104HalfEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c108BFloat16EEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c107complexIdEEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c107complexIfEEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c107complexINS2_4HalfEEEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c1011Float8_e5m2EEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c1013Float8_e4m3fnEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c1014Float8_e8m0fnuEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c1015Float8_e4m3fnuzEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c1015Float8_e5m2fnuzEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c105qint8EEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c106qint32EEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase16mutable_data_ptrIN3c106quint8EEEPT_v)
+
+// ---- data_ptr<T>  (legacy mutable accessor) ----
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIaEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIbEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIdEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIfEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIhEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIiEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIlEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIsEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIN3c104HalfEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIN3c108BFloat16EEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIN3c107complexIdEEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIN3c107complexIfEEEEPT_v)
+BRIDGE_WRITE(_ZNK2at10TensorBase8data_ptrIN3c107complexINS2_4HalfEEEEEPT_v)
+
+#undef BRIDGE_READ
+#undef BRIDGE_WRITE
+
+} // extern "C"
+
+#pragma GCC visibility pop
+#pragma GCC diagnostic pop
+// === openRuyi ROCm symbol bridge end ===
@@ -0,0 +1,44 @@
+diff --git a/setup.py b/setup.py
+index fe4a78d..788dc1c 100644
+--- a/setup.py
+++ b/setup.py
+@@ -463,7 +463,14 @@
+             "-DCMAKE_EXPORT_COMPILE_COMMANDS=ON", "-DLLVM_ENABLE_WERROR=ON",
+             "-DCMAKE_LIBRARY_OUTPUT_DIRECTORY=" + extdir, "-DTRITON_BUILD_PYTHON_MODULE=ON",
+             "-DPython3_EXECUTABLE:FILEPATH=" + sys.executable, "-DPython3_INCLUDE_DIR=" + python_include_dir,
+-            "-DTRITON_CODEGEN_BACKENDS=" + ';'.join([b.name for b in backends if not b.is_external]),
+            # openRuyi: core Triton hard-depends on the NVGPU/NVWS dialects that
+            # live under third_party/nvidia (TritonGPUTransforms and
+            # TritonInstrumentToLLVM include their TableGen output and link
+            # NVGPUIR/NVWSIR), so the nvidia backend must stay in the CMake
+            # build even though its Python side is not packaged (see the
+            # `backends` list below).
+            "-DTRITON_CODEGEN_BACKENDS=" +
+            ';'.join([b.name for b in backends if not b.is_external] + ["nvidia"]),
+             "-DTRITON_PLUGIN_DIRS=" + ';'.join([b.src_dir for b in backends if b.is_external]),
+             "-DTRITON_WHEEL_DIR=" + wheeldir
+         ]
+@@ -534,6 +541,10 @@
+ 
+ 
+ def download_and_copy_dependencies():
+    # openRuyi: this package ships only the AMD/ROCm backend, so the NVIDIA
+    # CUDA toolchain (ptxas, cuobjdump, ...) is neither needed nor downloaded.
+    # Skipping this also keeps the build fully offline for the OBS sandbox.
+    return
+     nvidia_version_path = os.path.join(get_base_dir(), "cmake", "nvidia-toolchain-version.json")
+     with open(nvidia_version_path, "r") as nvidia_version_file:
+         # parse this json file to get the version of the nvidia toolchain
+@@ -619,7 +630,11 @@
+     )
+ 
+ 
+-backends = [*BackendInstaller.copy(["nvidia", "amd"]), *BackendInstaller.copy_externals()]
+# openRuyi: ship the AMD/ROCm backend only.  The NVIDIA C++ libraries are
+# still compiled into libtriton (core requires them; see the cmake_args note
+# above), but the NVIDIA Python backend -- which would bundle ptxas and the
+# proprietary libdevice.10.bc -- is intentionally not packaged.
+backends = [*BackendInstaller.copy(["amd"]), *BackendInstaller.copy_externals()]
+ 
+ 
+ def get_package_dirs():
@@ -0,0 +1,21 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index c9620e3..6c1fbb0 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -262,6 +262,16 @@ if(TRITON_BUILD_PYTHON_MODULE)
+         LLVMPowerPCAsmParser
+         LLVMPowerPCCodeGen
+       )
+  elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "riscv64")
+      # There is no LLVM RISC-V GPU target; Triton never emits host code on
+      # riscv64.  However llvm::InitializeAllTargets() (referenced from
+      # llvm.cc) pulls in the X86 codegen symbols, so link them here to avoid
+      # an "undefined symbol: LLVMInitializeX86Target" failure at import time.
+      # The matching bundled LLVM must therefore be built with the X86 target.
+      list(APPEND TRITON_LIBRARIES
+          LLVMX86CodeGen
+          LLVMX86AsmParser
+      )
+   else()
+     message(FATAL_ERROR "LLVM codegen/ASM parser libs: This HW architecture (${CMAKE_SYSTEM_PROCESSOR}) is not configured in cmake lib dependencies.")
+   endif()
@@ -0,0 +1,25 @@
+diff --git a/setup.py b/setup.py
+index fe4a78d..400c34f 100644
+--- a/setup.py
+++ b/setup.py
+@@ -424,10 +424,19 @@ class CMakeBuild(build_ext):
+     def get_pybind11_cmake_args(self):
+         pybind11_sys_path = get_env_with_keys(["PYBIND11_SYSPATH"])
+         if pybind11_sys_path:
+            # openRuyi: distro pybind11 packages install the headers and the
+            # CMake config under a filesystem prefix (include/pybind11 and
+            # share/cmake/pybind11 below /usr), not inside the Python package
+            # the way pip wheels do, so pybind11.get_cmake_dir() raises
+            # ImportError ("pybind11 not installed").  When PYBIND11_SYSPATH
+            # is given, derive the CMake dir from it as well instead of only
+            # the include dir.
+             pybind11_include_dir = os.path.join(pybind11_sys_path, "include")
+            pybind11_cmake_dir = os.path.join(pybind11_sys_path, "share", "cmake", "pybind11")
+         else:
+             pybind11_include_dir = pybind11.get_include()
+-        return [f"-Dpybind11_INCLUDE_DIR='{pybind11_include_dir}'", f"-Dpybind11_DIR='{pybind11.get_cmake_dir()}'"]
+            pybind11_cmake_dir = pybind11.get_cmake_dir()
+        return [f"-Dpybind11_INCLUDE_DIR='{pybind11_include_dir}'", f"-Dpybind11_DIR='{pybind11_cmake_dir}'"]
+ 
+     def get_proton_cmake_args(self):
+         cmake_args = get_thirdparty_packages([get_json_package_info()])
@@ -0,0 +1,210 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+#
+# Originally extracted from Fedora Project
+# Authors: The Fedora Project Contributors
+# riscv64 build hints contributed by the openRuyi AI working group.
+
+%global srcname triton
+
+# Triton pins an exact, in-development LLVM *commit* (not a release version).
+# It calls unstable MLIR/LLVM C++ internals, so it only builds against that one
+# revision; no released distro LLVM (nor ROCm's bundled LLVM) matches it, and it
+# additionally needs MLIR and LLD.  We therefore build LLVM from source at the
+# pinned commit and link it statically into the Triton extension, exactly like
+# upstream's CI does.
+#
+# !!! WHEN BUMPING %%{version} !!!
+# Triton and LLVM must move together.  Set %%{llvm_commit} to the value of
+# cmake/llvm-hash.txt for the new Triton tag and refresh Source1's sha256.  A
+# mismatched LLVM will fail to compile or crash at runtime.
+%global llvm_commit f6ded0be897e2878612dd903f7e8bb85448269e5
+
+# Build everything (the bundled LLVM and the Triton extension) with clang,
+# matching the rest of the openRuyi ROCm stack.
+%global toolchain clang
+
+# The bundled static LLVM is large; drop LTO and skip the dwz pass which can
+# exhaust memory on the giant libtriton.so.
+%global _lto_cflags %{nil}
+%define _find_debuginfo_dwz_opts %{nil}
+
+Name:           python-%{srcname}
+Version:        3.6.0
+Release:        %autorelease
+Summary:        A language and compiler for custom Deep Learning operations
+# Triton itself is MIT.  The statically bundled LLVM/MLIR/LLD is
+# "Apache-2.0 WITH LLVM-exception OR NCSA"; pybind11 headers are BSD-3-Clause.
+License:        MIT AND (Apache-2.0 WITH LLVM-exception OR NCSA) AND BSD-3-Clause
+URL:            https://github.com/triton-lang/triton
+
+# Triton's PyPI sdist does not ship the C++ / third_party sources needed to
+# build, so the source is taken from the GitHub release tag instead.
+#!RemoteAsset:  sha256:be270ed11ca5a8fbd9d7941c5bbe9a23a9f6e2ffd372c8398346928bee464774
+Source0:        %{url}/archive/refs/tags/v%{version}.tar.gz#/%{srcname}-%{version}.tar.gz
+# NOTE: codeload generates llvm-project's commit archive on the fly; the
+# github.com/.../archive redirect to it times out behind the build proxy, so
+# point straight at codeload (identical bytes, same sha256).
+#!RemoteAsset:  sha256:f63c624aa63eda73508b9df2be2a6945ea4fddbee58615fbe1cd747b6884dd5e
+Source1:        https://github.com/llvm/llvm-project/archive/%{llvm_commit}.tar.gz
+
+# Ship only the AMD/ROCm Python backend and never reach out to the network
+# for the NVIDIA CUDA toolchain (ptxas, libdevice, ...).  The NVIDIA C++
+# libraries are still compiled into libtriton: Triton core hard-depends on
+# the NVGPU/NVWS dialects living under third_party/nvidia.
+Patch0:         0001-Ship-only-the-AMD-ROCm-backend-offline.patch
+# Link the X86 codegen libraries on the riscv64 host so that
+# llvm::InitializeAllTargets() resolves at import time.
+Patch1:         0002-Add-riscv64-host-codegen-libraries.patch
+# pybind11.get_cmake_dir() only knows the pip-wheel layout and raises
+# ImportError with a distro python3-pybind11, so let PYBIND11_SYSPATH (set in
+# %%build) supply the CMake dir as well.  The unconditional call came in with
+# https://github.com/triton-lang/triton/pull/4450
+Patch2:         0003-Use-PYBIND11_SYSPATH-for-the-pybind11-CMake-dir-too.patch
+BuildSystem:    pyproject
+
+BuildOption(install):  %{srcname}
+
+# --- Python build backend --------------------------------------------------
+BuildRequires:  pyproject-rpm-macros
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  python3dist(pip)
+BuildRequires:  python3dist(setuptools)
+BuildRequires:  python3dist(wheel)
+BuildRequires:  python3dist(pybind11)
+# Supplies %%{_includedir}/pybind11 and %%{_datadir}/cmake/pybind11, which
+# PYBIND11_SYSPATH points the build at (see Patch2).
+BuildRequires:  pkgconfig(pybind11)
+
+# --- Toolchain for the bundled LLVM and the Triton extension ---------------
+BuildRequires:  clang
+BuildRequires:  lld
+BuildRequires:  libstdc++-devel
+BuildRequires:  compiler-rt
+BuildRequires:  cmake
+BuildRequires:  ninja
+
+# --- Libraries the bundled LLVM links against ------------------------------
+BuildRequires:  pkgconfig(libffi)
+BuildRequires:  pkgconfig(libxml-2.0)
+BuildRequires:  pkgconfig(zlib)
+BuildRequires:  pkgconfig(libzstd)
+
+# --- Runtime ROCm stack ----------------------------------------------------
+# Triton JIT-compiles kernels at runtime: the GPU path uses the statically
+# linked LLD + AMDGPU code generator, but the per-kernel CPU launcher shim is
+# compiled on the fly (triton/runtime/build.py) with a host C compiler against
+# the Python headers, and the HIP runtime is dlopen'd.
+Requires:       gcc
+Requires:       pkgconfig(python3)
+Requires:       cmake(hip)
+Requires:       rocm-device-libs
+
+Provides:       python3-%{srcname} = %{version}-%{release}
+Provides:       python3-%{srcname}%{?_isa} = %{version}-%{release}
+%python_provide python3-%{srcname}
+
+%description
+Triton is a language and compiler for writing highly efficient custom
+Deep-Learning primitives. The aim of Triton is to provide an open-source
+environment to write fast code at higher productivity than CUDA, but also
+with higher flexibility than other existing DSLs.
+
+This build ships the AMD ROCm (HIP) backend.
+
+%prep -a
+# Unpack the pinned LLVM next to the Triton tree (built in %%build).
+tar -xf %{SOURCE1}
+
+# Drop any pre-generated metadata shipped in the tarball.
+rm -rf %{srcname}.egg-info
+
+# Triton's CMake turns warnings into errors; a from-source LLVM occasionally
+# emits new warnings, so relax this for both Triton and the embedded
+# add_llvm/add_mlir targets.
+sed -i -e 's@ -Werror @ @' CMakeLists.txt
+
+# The wheel is built with --no-build-isolation, so cmake/ninja/pybind11 are
+# supplied as system BuildRequires.  Strip them from build-system.requires so
+# %%pyproject_buildrequires does not emit unsatisfiable python3dist(cmake<4),
+# python3dist(ninja) dependencies.
+sed -i -e 's@^requires = .*@requires = ["setuptools>=40.8.0", "wheel"]@' pyproject.toml
+
+%generate_buildrequires
+%pyproject_buildrequires
+
+# Build the pinned LLVM+MLIR+LLD first, then let the pyproject build system
+# compile the Triton wheel against it.  Both run in the same shell, so the
+# environment exported here reaches %%pyproject_wheel.
+%build -p
+llvm_src="$(pwd)/llvm-project-%{llvm_commit}"
+llvm_install="$(pwd)/llvm-install"
+
+# Cap parallelism by available memory: LLVM/MLIR compile units and the final
+# Triton link are memory hungry and will thrash or OOM otherwise.
+mem_gb=$(awk '/MemTotal/ {print int($2/1024/1024)}' /proc/meminfo)
+compile_jobs=$(nproc)
+mem_jobs=$(( 1 + mem_gb / 2 ))
+[ "$mem_jobs" -lt "$compile_jobs" ] && compile_jobs=$mem_jobs
+[ "$compile_jobs" -lt 1 ] && compile_jobs=1
+# Linking the static archives needs far more memory per job.
+link_jobs=$(( 1 + mem_gb / 16 ))
+[ "$link_jobs" -lt 1 ] && link_jobs=1
+
+%ifarch x86_64
+llvm_targets="X86;AMDGPU;NVPTX"
+%endif
+%ifarch riscv64
+# X86 is required by the riscv64 codegen-libs patch; AMDGPU drives the ROCm
+# backend; NVPTX is always linked by Triton's core; RISCV is the host.
+llvm_targets="RISCV;X86;AMDGPU;NVPTX"
+%endif
+
+cmake -S "$llvm_src/llvm" -B "$llvm_src/build" -G Ninja \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_INSTALL_PREFIX="$llvm_install" \
+    -DCMAKE_C_COMPILER=clang \
+    -DCMAKE_CXX_COMPILER=clang++ \
+    -DLLVM_USE_LINKER=lld \
+    -DLLVM_ENABLE_PROJECTS="mlir;lld" \
+    -DLLVM_TARGETS_TO_BUILD="$llvm_targets" \
+    -DLLVM_ENABLE_ASSERTIONS=OFF \
+    -DBUILD_SHARED_LIBS=OFF \
+    -DLLVM_BUILD_LLVM_DYLIB=OFF \
+    -DLLVM_INSTALL_UTILS=ON \
+    -DLLVM_ENABLE_TERMINFO=OFF \
+    -DLLVM_ENABLE_ZSTD=ON \
+    -DLLVM_INCLUDE_BENCHMARKS=OFF \
+    -DLLVM_INCLUDE_EXAMPLES=OFF \
+    -DLLVM_INCLUDE_TESTS=OFF \
+    -DLLVM_PARALLEL_COMPILE_JOBS=$compile_jobs \
+    -DLLVM_PARALLEL_LINK_JOBS=$link_jobs
+cmake --build "$llvm_src/build" --target install -- -j$compile_jobs
+
+# Point Triton at the freshly built LLVM and keep the build offline + ROCm-only.
+export LLVM_SYSPATH="$llvm_install"
+export PATH="$llvm_install/bin:$PATH"
+# System pybind11 from pybind11-devel: headers and CMake config under
+# %%{_prefix} (see Patch2).
+export PYBIND11_SYSPATH=%{_prefix}
+export CC=clang
+export CXX=clang++
+export MAX_JOBS=$compile_jobs
+export TRITON_PARALLEL_LINK_JOBS=$link_jobs
+export TRITON_BUILD_WITH_CLANG_LLD=ON
+export TRITON_BUILD_WITH_CCACHE=OFF
+# Proton needs CUPTI/roctracer/json; not needed for a plain ROCm backend.
+export TRITON_BUILD_PROTON=OFF
+# Don't fetch googletest, and don't trip over new LLVM warnings.
+export TRITON_APPEND_CMAKE_ARGS="-DTRITON_BUILD_UT=OFF -DLLVM_ENABLE_WERROR=OFF"
+
+%files -f %{pyproject_files}
+# Triton's own LICENSE is captured from the wheel metadata by
+# %%pyproject_save_files; only the bundled LLVM license must be added by hand.
+%license llvm-project-%{llvm_commit}/llvm/LICENSE.TXT
+%doc README.md
+
+%changelog
+%autochangelog
@@ -0,0 +1,133 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           rccl
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm Communication Collectives Library
+License:        BSD-3-Clause AND MIT AND Apache-2.0
+# From License.txt the main license is BSD 3
+# Modifications from Microsoft is MIT
+# The NVIDIA based header files below are Apache-2.0
+#  src/include/nvtx3/nv*.h and similar
+# The URL for NVIDIA in the License.txt https://github.com/NVIDIA/NVTX is Apache-2.0
+Url:            https://github.com/ROCm/rccl
+#!RemoteAsset:  sha256:eaa60bcf62feb3198553f2bcf6dcbfdfcecd0fdfabda41f1dae7d3f15fadbd68
+Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DEXPLICIT_ROCM_VERSION=%{rocm_version}
+BuildOption(conf):  -DROCM_PATH=%{_prefix}
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+BuildOption(conf):  -DBUILD_TESTS=ON
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(fmt)
+BuildRequires:  cmake(GTest)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(rocm_smi)
+BuildRequires:  cmake(rocm-core)
+BuildRequires:  cmake(rocprofiler-register)
+BuildRequires:  compiler-rt
+BuildRequires:  hipify
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  python3
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-llvm-macros
+
+Requires:       %{name}-data = %{version}-%{release}
+
+%description
+RCCL (pronounced "Rickle") is a stand-alone library of standard
+collective communication routines for GPUs, implementing all-reduce,
+all-gather, reduce, broadcast, reduce-scatter, gather, scatter, and
+all-to-all. There is also initial support for direct GPU-to-GPU
+send and receive operations. It has been optimized to achieve high
+bandwidth on platforms using PCIe, xGMI as well as networking using
+InfiniBand Verbs or TCP/IP sockets. RCCL supports an arbitrary
+number of GPUs installed in a single node or multiple nodes, and
+can be used in either single- or multi-process (e.g., MPI)
+applications.
+
+The collective operations are implemented using ring and tree
+algorithms and have been optimized for throughput and latency. For
+best performance, small operations can be either batched into
+larger operations or aggregated through the API.
+
+%package        devel
+Summary:        Headers and libraries for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Provides:       rccl-devel = %{version}-%{release}
+
+%description    devel
+Headers and libraries for %{name}
+
+%package        data
+Summary:        Data for %{name}
+BuildArch:      noarch
+
+%description    data
+Data for %{name}
+
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+
+%prep -a
+# Do not force install
+sed -i -e 's@set(CMAKE_INSTALL_LIBDIR@#set(CMAKE_INSTALL_LIBDIR@' cmake/Dependencies.cmake
+# -amdgpu-s-branch-bits and -amdgpu-long-branch-factor=2 are needed to avoid 'branch size exceed simm16' error
+# --lto-partitions to accelerate linking time
+sed -i -e 's@target_link_options(rccl PRIVATE "SHELL:-Xoffload-linker -mllvm=-amdgpu-kernarg-preload-count=16")@target_link_options(rccl PRIVATE "SHELL:-Xoffload-linker -mllvm=-amdgpu-s-branch-bits=15" "SHELL:-Xoffload-linker -mllvm=-amdgpu-long-branch-factor=2" "SHELL:-Xoffload-linker -mllvm=-amdgpu-kernarg-preload-count=16" "SHELL:-Xoffload-linker --lto-partitions=%(nproc)" "SHELL:-Xoffload-linker --verbose")@' CMakeLists.txt
+
+%build
+# AMDGPU device linker runs as a process that produces no stdout for about 8~12 hours on riscv64
+timeout 12h bash -c 'while sleep 300; do echo "[heartbeat] $(date)"; done' & TIME_OUT=$!
+%cmake_build
+kill $TIME_OUT 2>/dev/null || true
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/rccl/LICENSE.txt
+
+%files
+%license LICENSE.txt
+%{_libdir}/librccl.so.*
+%{_bindir}/rcclras
+
+%files data
+%{_datadir}/rccl/msccl-algorithms/
+%{_datadir}/rccl/msccl-unit-test-algorithms/
+
+%files devel
+%doc README.md
+%{_includedir}/rccl/
+%{_libdir}/cmake/rccl/
+%{_libdir}/librccl.so
+
+%files test
+%{_bindir}/rccl-UnitTests
+
+%changelog
+%autochangelog
@@ -1,25 +0,0 @@
-From 87f3c0b3ebab78aa6d126633683720031d886313 Mon Sep 17 00:00:00 2001
-From: Tom Rix <trix@redhat.com>
-Date: Sat, 13 Jan 2024 14:36:01 -0500
-Subject: [PATCH] fixup install of tensile output
-
---
- library/src/CMakeLists.txt | 2 +-
- 1 file changed, 1 insertion(+), 1 deletion(-)
-
-diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
-index f4bdfb5f9742..316779134314 100644
--- a/library/src/CMakeLists.txt
-+++ b/library/src/CMakeLists.txt
-@@ -823,7 +823,7 @@ if( BUILD_WITH_TENSILE )
-   if (WIN32)
-     set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}/bin/rocblas" CACHE PATH "path to tensile library" )
-   else()
-    set( ROCBLAS_TENSILE_LIBRARY_DIR "\${CPACK_PACKAGING_INSTALL_PREFIX}${CMAKE_INSTALL_LIBDIR}/rocblas" CACHE PATH "path to tensile library" )
-+    set( ROCBLAS_TENSILE_LIBRARY_DIR "${CMAKE_INSTALL_LIBDIR}/rocblas" CACHE PATH "path to tensile library" )
-   endif()
-   # For ASAN package, Tensile library files(which are not shared libraries) are not required
-   if( NOT ENABLE_ASAN_PACKAGING )
-- 
-2.51.0
-
@@ -1,314 +0,0 @@
-diff --git a/clients/include/blas2/testing_gbmv.hpp b/clients/include/blas2/testing_gbmv.hpp
-index d02e1a5..bfe1046 100644
--- a/clients/include/blas2/testing_gbmv.hpp
-+++ b/clients/include/blas2/testing_gbmv.hpp
-@@ -267,11 +267,11 @@ void testing_gbmv(const Arguments& arg)
-     hy_gold = hy;
- 
-     // copy data from CPU to device
-    dAb.transfer_from(hAb);
-    dx.transfer_from(hx);
-    dy.transfer_from(hy);
-    d_alpha.transfer_from(halpha);
-    d_beta.transfer_from(hbeta);
-+    CHECK_HIP_ERROR(dAb.transfer_from(hAb));
-+    CHECK_HIP_ERROR(dx.transfer_from(hx));
-+    CHECK_HIP_ERROR(dy.transfer_from(hy));
-+    CHECK_HIP_ERROR(d_alpha.transfer_from(halpha));
-+    CHECK_HIP_ERROR(d_beta.transfer_from(hbeta));
- 
-     double cpu_time_used;
-     double error_host = 0.0, error_device = 0.0;
-@@ -290,12 +290,12 @@ void testing_gbmv(const Arguments& arg)
-                 (handle, transA, M, N, KL, KU, &h_alpha, dAb, lda, dx, incx, &h_beta, dy, incy));
-             handle.post_test(arg);
- 
-            hy.transfer_from(dy);
-+            CHECK_HIP_ERROR(hy.transfer_from(dy));
-         }
- 
-         if(arg.pointer_mode_device)
-         {
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             CHECK_ROCBLAS_ERROR(rocblas_set_pointer_mode(handle, rocblas_pointer_mode_device));
-             handle.pre_test(arg);
-@@ -308,7 +308,7 @@ void testing_gbmv(const Arguments& arg)
-             {
-                 HOST_MEMCHECK(host_vector<T>, hy_copy, (dim_y, incy));
-                 // copy output from device to CPU
-                hy.transfer_from(dy);
-+                CHECK_HIP_ERROR(hy.transfer_from(dy));
- 
-                 // multi-GPU support
-                 int device_id, device_count;
-@@ -330,17 +330,17 @@ void testing_gbmv(const Arguments& arg)
-                     DEVICE_MEMCHECK(device_vector<T>, d_alpha_copy, (1));
-                     DEVICE_MEMCHECK(device_vector<T>, d_beta_copy, (1));
- 
-                    dAb_copy.transfer_from(hAb);
-                    dx_copy.transfer_from(hx);
-                    d_alpha_copy.transfer_from(halpha);
-                    d_beta_copy.transfer_from(hbeta);
-+                    CHECK_HIP_ERROR(dAb_copy.transfer_from(hAb));
-+                    CHECK_HIP_ERROR(dx_copy.transfer_from(hx));
-+                    CHECK_HIP_ERROR(d_alpha_copy.transfer_from(halpha));
-+                    CHECK_HIP_ERROR(d_beta_copy.transfer_from(hbeta));
- 
-                     CHECK_ROCBLAS_ERROR(
-                         rocblas_set_pointer_mode(handle_copy, rocblas_pointer_mode_device));
- 
-                     for(int runs = 0; runs < arg.iters; runs++)
-                     {
-                        dy_copy.transfer_from(hy_gold);
-+                        CHECK_HIP_ERROR(dy_copy.transfer_from(hy_gold));
-                         DAPI_CHECK(rocblas_gbmv_fn,
-                                    (handle_copy,
-                                     transA,
-@@ -357,7 +357,7 @@ void testing_gbmv(const Arguments& arg)
-                                     dy_copy,
-                                     incy));
-                         // copy output from device to CPU
-                        hy_copy.transfer_from(dy_copy);
-+                        CHECK_HIP_ERROR(hy_copy.transfer_from(dy_copy));
-                         unit_check_general<T>(1, dim_y, incy, hy, hy_copy);
-                     }
-                 }
-@@ -383,7 +383,7 @@ void testing_gbmv(const Arguments& arg)
-         if(arg.pointer_mode_device)
-         {
-             // copy output from device to CPU
-            hy.transfer_from(dy);
-+            CHECK_HIP_ERROR(hy.transfer_from(dy));
- 
-             if(arg.unit_check)
-             {
-diff --git a/clients/include/blas2/testing_sbmv.hpp b/clients/include/blas2/testing_sbmv.hpp
-index feb1148..95e09e2 100644
--- a/clients/include/blas2/testing_sbmv.hpp
-+++ b/clients/include/blas2/testing_sbmv.hpp
-@@ -204,9 +204,9 @@ void testing_sbmv(const Arguments& arg)
-     hy_gold = hy;
- 
-     // copy data from CPU to device
-    dx.transfer_from(hx);
-    dy.transfer_from(hy);
-    dAb.transfer_from(hAb);
-+    CHECK_HIP_ERROR(dx.transfer_from(hx));
-+    CHECK_HIP_ERROR(dy.transfer_from(hy));
-+    CHECK_HIP_ERROR(dAb.transfer_from(hAb));
- 
-     double cpu_time_used;
-     double error_host = 0.0, error_device = 0.0;
-@@ -231,7 +231,7 @@ void testing_sbmv(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
- 
-diff --git a/clients/include/blas2/testing_sbmv_batched.hpp b/clients/include/blas2/testing_sbmv_batched.hpp
-index 4812be9..9a91392 100644
--- a/clients/include/blas2/testing_sbmv_batched.hpp
-+++ b/clients/include/blas2/testing_sbmv_batched.hpp
-@@ -322,9 +322,9 @@ void testing_sbmv_batched(const Arguments& arg)
-     hy_gold.copy_from(hy);
- 
-     // copy data from CPU to device
-    dx.transfer_from(hx);
-    dy.transfer_from(hy);
-    dAb.transfer_from(hAb);
-+    CHECK_HIP_ERROR(dx.transfer_from(hx));
-+    CHECK_HIP_ERROR(dy.transfer_from(hy));
-+    CHECK_HIP_ERROR(dAb.transfer_from(hAb));
- 
-     double cpu_time_used;
-     double h_error = 0.0, d_error = 0.0;
-@@ -363,7 +363,7 @@ void testing_sbmv_batched(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
-             DAPI_CHECK(rocblas_sbmv_batched_fn,
-diff --git a/clients/include/blas2/testing_sbmv_strided_batched.hpp b/clients/include/blas2/testing_sbmv_strided_batched.hpp
-index a32538a..ad902a3 100644
--- a/clients/include/blas2/testing_sbmv_strided_batched.hpp
-+++ b/clients/include/blas2/testing_sbmv_strided_batched.hpp
-@@ -385,9 +385,9 @@ void testing_sbmv_strided_batched(const Arguments& arg)
-     hy_gold.copy_from(hy);
- 
-     // copy data from CPU to device
-    dx.transfer_from(hx);
-    dy.transfer_from(hy);
-    dAb.transfer_from(hAb);
-+    CHECK_HIP_ERROR(dx.transfer_from(hx));
-+    CHECK_HIP_ERROR(dy.transfer_from(hy));
-+    CHECK_HIP_ERROR(dAb.transfer_from(hAb));
- 
-     double cpu_time_used;
-     double error_host = 0.0, error_device = 0.0;
-@@ -428,7 +428,7 @@ void testing_sbmv_strided_batched(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
-             DAPI_CHECK(rocblas_sbmv_strided_batched_fn,
-diff --git a/clients/include/blas2/testing_symv.hpp b/clients/include/blas2/testing_symv.hpp
-index 2b31355..d170478 100644
--- a/clients/include/blas2/testing_symv.hpp
-+++ b/clients/include/blas2/testing_symv.hpp
-@@ -213,7 +213,7 @@ void testing_symv(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
-             DAPI_CHECK(rocblas_symv_fn,
-diff --git a/clients/include/blas2/testing_symv_batched.hpp b/clients/include/blas2/testing_symv_batched.hpp
-index 6fd3f7b..ceed6f3 100644
--- a/clients/include/blas2/testing_symv_batched.hpp
-+++ b/clients/include/blas2/testing_symv_batched.hpp
-@@ -345,7 +345,7 @@ void testing_symv_batched(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
-             DAPI_CHECK(rocblas_symv_batched_fn,
-diff --git a/clients/include/blas2/testing_symv_strided_batched.hpp b/clients/include/blas2/testing_symv_strided_batched.hpp
-index d96e17c..dbaa454 100644
--- a/clients/include/blas2/testing_symv_strided_batched.hpp
-+++ b/clients/include/blas2/testing_symv_strided_batched.hpp
-@@ -433,7 +433,7 @@ void testing_symv_strided_batched(const Arguments& arg)
-             CHECK_HIP_ERROR(d_alpha.transfer_from(alpha));
-             CHECK_HIP_ERROR(d_beta.transfer_from(beta));
- 
-            dy.transfer_from(hy_gold);
-+            CHECK_HIP_ERROR(dy.transfer_from(hy_gold));
- 
-             handle.pre_test(arg);
-             DAPI_CHECK(rocblas_symv_strided_batched_fn,
-diff --git a/clients/include/blas_ex/testing_gemm_batched_ex.hpp b/clients/include/blas_ex/testing_gemm_batched_ex.hpp
-index 214f0b4..54ca0b5 100644
--- a/clients/include/blas_ex/testing_gemm_batched_ex.hpp
-+++ b/clients/include/blas_ex/testing_gemm_batched_ex.hpp
-@@ -103,7 +103,7 @@ void testing_gemm_batched_ex_bad_arg(const Arguments& arg)
-         rocblas_seedrand();
-         rocblas_init_matrix<To>(
-             hC, arg, rocblas_client_beta_sets_nan, rocblas_client_general_matrix);
-        dC.transfer_from(hC);
-+        CHECK_HIP_ERROR(dC.transfer_from(hC));
- 
-         // clang-format off
- // check for invalid enum
-diff --git a/clients/include/blas_ex/testing_gemm_ex.hpp b/clients/include/blas_ex/testing_gemm_ex.hpp
-index 4977995..be36f2e 100644
--- a/clients/include/blas_ex/testing_gemm_ex.hpp
-+++ b/clients/include/blas_ex/testing_gemm_ex.hpp
-@@ -102,7 +102,7 @@ void testing_gemm_ex_bad_arg(const Arguments& arg)
-         HOST_MEMCHECK(host_matrix<To>, hC, (M, N, ldc));
-         rocblas_seedrand();
-         rocblas_init_matrix(hC, arg, rocblas_client_beta_sets_nan, rocblas_client_general_matrix);
-        dC.transfer_from(hC);
-+        CHECK_HIP_ERROR(dC.transfer_from(hC));
- 
-         // clang-format off
- 
-diff --git a/clients/include/testing_set_get_matrix_async.hpp b/clients/include/testing_set_get_matrix_async.hpp
-index 01d0648..de8d301 100644
--- a/clients/include/testing_set_get_matrix_async.hpp
-+++ b/clients/include/testing_set_get_matrix_async.hpp
-@@ -127,7 +127,7 @@ void testing_set_get_matrix_async(const Arguments& arg)
- 
-         cpu_time_used = get_time_us_no_sync() - cpu_time_used;
- 
-        hipStreamSynchronize(stream);
-+        CHECK_HIP_ERROR(hipStreamSynchronize(stream));
- 
-         if(arg.unit_check)
-         {
-@@ -160,7 +160,7 @@ void testing_set_get_matrix_async(const Arguments& arg)
-                           (rows, cols, sizeof(T), dD, ldd, hB, ldb, stream));
-         }
- 
-        hipStreamSynchronize(stream);
-+        CHECK_HIP_ERROR(hipStreamSynchronize(stream));
-         gpu_time_used = get_time_us_sync(stream) - gpu_time_used;
- 
-         ArgumentModel<e_M, e_N, e_lda, e_ldb, e_ldd>{}.log_args<T>(
-diff --git a/clients/include/testing_set_get_vector_async.hpp b/clients/include/testing_set_get_vector_async.hpp
-index b88cc0c..0bddc20 100644
--- a/clients/include/testing_set_get_vector_async.hpp
-+++ b/clients/include/testing_set_get_vector_async.hpp
-@@ -113,7 +113,7 @@ void testing_set_get_vector_async(const Arguments& arg)
- 
-         cpu_time_used = get_time_us_no_sync() - cpu_time_used;
- 
-        hipStreamSynchronize(stream);
-+        CHECK_HIP_ERROR(hipStreamSynchronize(stream));
- 
-         if(arg.unit_check)
-         {
-@@ -144,7 +144,7 @@ void testing_set_get_vector_async(const Arguments& arg)
-             DAPI_DISPATCH(rocblas_get_vector_async_fn, (N, sizeof(T), db, ldd, hy, incy, stream));
-         }
- 
-        hipStreamSynchronize(stream);
-+        CHECK_HIP_ERROR(hipStreamSynchronize(stream));
-         gpu_time_used = get_time_us_sync(stream) - gpu_time_used;
- 
-         ArgumentModel<e_N, e_incx, e_incy, e_ldd>{}.log_args<T>(rocblas_cout,
-diff --git a/library/src/include/handle.hpp b/library/src/include/handle.hpp
-index a0a1760..c80cc5e 100644
--- a/library/src/include/handle.hpp
-+++ b/library/src/include/handle.hpp
-@@ -147,16 +147,20 @@ private:
-             : device_id(device_id)
-             , old_device_id(-1)
-         {
-            hipGetDevice(&old_device_id);
-+            THROW_IF_HIP_ERROR(hipGetDevice(&old_device_id));
-             if(device_id != old_device_id)
-                hipSetDevice(device_id);
-+            {
-+                THROW_IF_HIP_ERROR(hipSetDevice(device_id));
-+            }
-         }
- 
-         // Old device ID is restored on destruction
-         ~_rocblas_saved_device_id()
-         {
-             if(device_id != old_device_id)
-                hipSetDevice(old_device_id);
-+            {
-+                (void)(hipSetDevice(old_device_id));
-+            }
-         }
- 
-         // Move constructor
-diff --git a/library/src/src64/blas1/rocblas_dot_kernels_64.cpp b/library/src/src64/blas1/rocblas_dot_kernels_64.cpp
-index 0bd3061..2b38dd7 100644
--- a/library/src/src64/blas1/rocblas_dot_kernels_64.cpp
-+++ b/library/src/src64/blas1/rocblas_dot_kernels_64.cpp
-@@ -325,7 +325,7 @@ rocblas_status rocblas_internal_dot_launcher_64(rocblas_handle __restrict__ hand
-     if(handle->pointer_mode == rocblas_pointer_mode_host)
-     {
-         // sync here to match legacy BLAS
-        hipStreamSynchronize(handle->get_stream());
-+        RETURN_IF_HIP_ERROR(hipStreamSynchronize(handle->get_stream()));
-     }
- 
-     return rocblas_status_success;
@@ -1,168 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-# Tests consumes too much time and space
-%bcond test 0
-
-%global rocm_version 7.1.1
-
-Name:           rocblas
-Summary:        BLAS implementation for ROCm
-Version:        %{rocm_version}
-Release:        %autorelease
-License:        MIT AND BSD-3-Clause
-URL:            https://github.com/ROCm/rocBLAS
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-BuildOption(conf):  -G Ninja
-BuildOption(conf):  -DBLAS_INCLUDE_DIR=%{_includedir}/cblas
-BuildOption(conf):  -DBLAS_LIBRARY=cblas
-BuildOption(conf):  -DCMAKE_CXX_COMPILER=hipcc
-BuildOption(conf):  -DCMAKE_C_COMPILER=clang
-BuildOption(conf):  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld
-BuildOption(conf):  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar
-BuildOption(conf):  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib
-BuildOption(conf):  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/..
-BuildOption(conf):  -DCMAKE_SKIP_RPATH=ON
-BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
-BuildOption(conf):  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
-# Avoid using external tensile
-BuildOption(conf):  -DBUILD_WITH_PIP=OFF
-BuildOption(conf):  -DROCM_SYMLINK_LIBS=OFF
-BuildOption(conf):  -DHIP_PLATFORM=amd
-# These will be enabled in a long future
-BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=%{?with_test:ON}%{!?with_test:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{?with_test:ON}%{!?with_test:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_TESTS_OPENMP=OFF
-BuildOption(conf):  -DBUILD_FORTRAN_CLIENTS=OFF
-BuildOption(conf):  -DBUILD_OFFLOAD_COMPRESS=ON
-BuildOption(conf):  -DBUILD_WITH_HIPBLASLT=OFF
-BuildOption(conf):  -DBUILD_WITH_TENSILE=ON
-BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
-BuildOption(conf):  -DTensile_LIBRARY_FORMAT=msgpack
-BuildOption(conf):  -DTensile_VERBOSE=1
-BuildOption(conf):  -DTensile_DIR=$(%{_bindir}/TensileGetPath)/cmake
-BuildOption(conf):  -DTensile_LOGIC=asm_full
-BuildOption(conf):  -DTensile_CODE_OBJECT_VERSION=default
-BuildOption(conf):  -DTensile_SEPARATE_ARCHITECTURES=ON
-BuildOption(conf):  -DTensile_LAZY_LIBRARY_LOADING=ON
-BuildOption(conf):  -DTensile_ASSEMBLER=clang++
-
-Patch0:         0001-fixup-install-of-tensile-output.patch
-# https://github.com/ROCm/rocm-libraries/commit/6221075881f3ea8e9dfa0d985f22005c74ae1f52
-Patch1:         0002-fix-nodiscard-return-value-ignored.patch
-
-BuildRequires:  clang
-BuildRequires:  clang-tools-extra
-BuildRequires:  cmake
-BuildRequires:  cmake(amd_comgr)
-BuildRequires:  cmake(hip)
-BuildRequires:  cmake(hsa-runtime64)
-BuildRequires:  cmake(msgpack)
-BuildRequires:  compiler-rt
-BuildRequires:  gcc-c++
-BuildRequires:  hipcc
-BuildRequires:  lld
-BuildRequires:  llvm
-BuildRequires:  ninja
-BuildRequires:  pkgconfig(libzstd)
-BuildRequires:  python3dist(tensile)
-BuildRequires:  rocm-cmake
-BuildRequires:  rocm-device-libs
-BuildRequires:  rocm-llvm-macros
-
-%if %{with test}
-BuildRequires:  gcc-fortran
-BuildRequires:  cmake(openmp)
-BuildRequires:  cmake(rocm_smi)
-BuildRequires:  pkgconfig(blas)
-BuildRequires:  pkgconfig(GTest)
-BuildRequires:  python3dist(pyyaml)
-BuildRequires:  rocminfo
-%endif
-
-Provides:       rocblas = %{version}-%{release}
-Requires:       python3dist(msgpack)
-
-%description
-rocBLAS is the AMD library for Basic Linear Algebra Subprograms
-(BLAS) on the ROCm platform. It is implemented in the HIP
-programming language and optimized for AMD GPUs.
-
-%package        devel
-Summary:        Libraries and headers for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-Requires:       rocm-hip-devel
-
-%description    devel
-%{summary}
-
-%if %{with test}
-%package        test
-Summary:        Tests for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-Requires:       diffutils
-
-%description    test
-%{summary}
-%endif
-
-%prep -a
-sed -i -e 's@target_link_libraries( rocblas-test PRIVATE ${BLAS_LIBRARY} ${GTEST_BOTH_LIBRARIES} roc::rocblas )@target_link_libraries( rocblas-test PRIVATE cblas ${GTEST_BOTH_LIBRARIES} roc::rocblas )@' clients/gtest/CMakeLists.txt
-
-# no git in this build
-sed -i -e 's@find_package(Git REQUIRED)@find_package(Git)@' library/CMakeLists.txt
-
-# /usr/include/gtest/internal/gtest-port.h:279:2: error: C++ versions less than C++14 are not supported.
-#   279 | #error C++ versions less than C++14 are not supported.
-sed -i -e 's@CXX_STANDARD 11@CXX_STANDARD 17@' clients/samples/CMakeLists.txt
-sed -i "s@/opt/rocm@%{_prefix}@g" \
-    clients/cmake/FindROCmSMI.cmake \
-    clients/CMakeLists.txt \
-    rmake.py \
-    rmake.py \
-    rmake.py \
-    toolchain-linux.cmake \
-    header_compilation_tests.sh \
-    library/src/tensile_host.cpp \
-    library/src/include/handle.hpp \
-    scripts/utilities/check_for_pretuned_sizes_c/Makefile \
-    scripts/performance/blas/getspecs.py \
-    scripts/performance/blas/commandrunner.py \
-    CMakeLists.txt \
-    library/CMakeLists.txt
-sed -i "s@llvm/bin@bin@g" CMakeLists.txt library/CMakeLists.txt
-
-%install -a
-rm -f %{buildroot}%{_prefix}/share/doc/rocblas/LICENSE.md
-
-%check
-%if %{with test}
-export LD_LIBRARY_PATH=%{_vpath_builddir}/library/src:$LD_LIBRARY_PATH
-%{_vpath_builddir}/clients/staging/rocblas-test --gtest_brief=1
-%endif
-
-%files
-%license LICENSE.md
-%{_libdir}/librocblas.so.5{,.*}
-%{_libdir}/rocblas/
-
-%files devel
-%doc README.md
-%{_includedir}/rocblas/
-%{_libdir}/cmake/rocblas/
-%{_libdir}/librocblas.so
-
-%if %{with test}
-%files test
-%{_bindir}/rocblas*
-%endif
-
-%changelog
-%{autochangelog}
@@ -0,0 +1,13 @@
+diff --git a/CMakeLists.txt b/CMakeLists.txt
+index 14bb20b..a5b1ee4 100644
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -49,6 +49,8 @@ set( ROCFFT_BUILD_SCOPE ON )
+ 
+ project( rocfft LANGUAGES CXX C )
+ 
+include(GNUInstallDirs)
+
+ # This finds the rocm-cmake project, and installs it if not found
+ # rocm-cmake contains common cmake code for rocm projects to help setup and install
+ set( PROJECT_EXTERN_DIR ${CMAKE_CURRENT_BINARY_DIR}/extern )
@@ -0,0 +1,111 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# Without a GPU, the test cases will fail with `what():  hipGetDeviceCount failed`
+# rocFFT needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU, so make it optional.
+%bcond test 0
+
+%global rocm_release 7.1
+%global rocm_patch 1 
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           rocfft
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm Fast Fourier Transforms library
+License:        MIT
+Url:            https://github.com/ROCm/rocFFT
+#!RemoteAsset:  sha256:047e4e93e0b12869bf42136b5eb683df3a1635b01a58bbb25c8861df291ab285
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+Patch0:         0001-cmake-use-gnu-installdirs.patch
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=ON
+BuildOption(conf):  -DROCFFT_BUILD_OFFLINE_TUNER=OFF
+BuildOption(conf):  -DROCFFT_KERNEL_CACHE_ENABLE=OFF
+BuildOption(conf):  -DSQLITE_USE_SYSTEM_PACKAGE=ON
+
+BuildRequires:  boost-devel
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hiprand)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  cmake(GTest)
+BuildRequires:  cmake(rocrand)
+BuildRequires:  compiler-rt
+BuildRequires:  libomp-devel
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(fftw3)
+BuildRequires:  pkgconfig(sqlite3)
+BuildRequires:  python3
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+rocFFT is a software library for computing fast Fourier transforms (FFTs) written
+in HIP. It is part of AMD's software ecosystem based on ROCm. In addition to
+AMD GPU hardware, rocFFT also works on CPU devices to facilitate testing.
+
+%package        devel
+Summary:        The rocFFT development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(hip)
+
+%description    devel
+The rocFFT development package.
+
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+
+%prep -a
+# Do not care so much about the sqlite version
+sed -i -e 's@SQLite3 3.50.2 @SQLite3 @' cmake/sqlite.cmake
+
+%install -a
+# we don't need the rocfft_rtc_helper binary and client-info file
+find %{buildroot} -type f -name "rocfft_rtc_helper" -print0 | xargs -0 -I {} /usr/bin/rm -rf "{}"
+rm -rf %{buildroot}/%{_prefix}/.info
+rm -f %{buildroot}%{_datadir}/doc/rocfft/LICENSE.md
+
+%if %{with test}
+%check
+%{_vpath_builddir}/clients/staging/rocfft-test
+%endif
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/librocfft.so.0{,.*}
+
+%files devel
+%{_includedir}/rocfft/
+%{_libdir}/cmake/rocfft/
+%{_libdir}/librocfft.so
+
+%files test
+%{_bindir}/rocfft-test
+%{_bindir}/rtc_helper_crash
+
+%changelog
+%autochangelog
@@ -0,0 +1,52 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: Sakura286 <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_version 7.1.1
+
+Name:           rocm-core
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        A utility to get the ROCm release version
+License:        MIT
+URL:            https://github.com/ROCm/rocm-core
+#!RemoteAsset:  sha256:0171b82a4d028d57035d0d57a01a058f50f1a23959d230cdeab14972dcd94da8
+Source0:        %{url}/archive/refs/tags/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -DROCM_VERSION=%{rocm_version}
+
+BuildRequires:  cmake
+
+Provides:       rocm-core = %{version}-%{release}
+
+%description
+%{summary}
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%install -a
+rm -rvf %{buildroot}/%{_exec_prefix}/.info
+rm -rvf %{buildroot}/%{_exec_prefix}/libexec/rocm-core
+rm -rvf %{buildroot}/%{_exec_prefix}/share/doc/*/LICENSE.md
+rm -rvf %{buildroot}/%{_libdir}/rocmmod
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/librocm-core.so.*
+
+%files devel
+%{_includedir}/rocm-core/*.h
+%{_libdir}/cmake/rocm-core/*.cmake
+%{_libdir}/librocm-core.so
+
+%changelog
+%{?autochangelog}
@@ -0,0 +1,15 @@
+diff --git a/amd/comgr/CMakeLists.txt b/amd/comgr/CMakeLists.txt
+index cfa170f94..e03049224 100644
+--- a/amd/comgr/CMakeLists.txt
+++ b/amd/comgr/CMakeLists.txt
+@@ -169,6 +169,10 @@ if (ADDRESS_SANITIZER)
+     "${CMAKE_SHARED_LINKER_FLAGS} ${ASAN_LINKER_FLAGS}")
+ endif()
+ 
+
+set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${ASAN_COMPILER_FLAGS} -fsigned-char")
+set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${ASAN_COMPILER_FLAGS} -fsigned-char")
+
+ set(AMD_COMGR_PRIVATE_COMPILE_OPTIONS)
+ set(AMD_COMGR_PRIVATE_COMPILE_DEFINITIONS ${LLVM_DEFINITIONS})
+ set(AMD_COMGR_PUBLIC_LINKER_OPTIONS)
@@ -0,0 +1,122 @@
+ROCm 7.1.1 build with llvm-20, but only llvm-21 is provided by openRuyi.
+Some backport work need to be done.
+
+This patch includes:
+
+* https://github.com/ROCm/llvm-project/commit/a7ad03285bf9ff361acb5e721386870be9354620
+* https://github.com/ROCm/llvm-project/commit/75074634076c0e3b8b2a18bbcf6ffef01094b069
+
+and another diff:
+
+old https://github.com/ROCm/llvm-project/commit/f2987311af76ab1d5e6f770861865b6952002b0a
+new https://github.com/ROCm/llvm-project/commit/e3dc0a41658572aecb595e55a79b9f4a85224187
+
+---
+diff --git a/amd/comgr/src/comgr-cache-bundler-command.cpp b/amd/comgr/src/comgr-cache-bundler-command.cpp
+index 514262725..23e919dbd 100644
+--- a/amd/comgr/src/comgr-cache-bundler-command.cpp
+++ b/amd/comgr/src/comgr-cache-bundler-command.cpp
+@@ -155,10 +155,8 @@ void UnbundleCommand::addOptionsIdentifier(HashAlgorithm &H) const {
+ Error UnbundleCommand::addInputIdentifier(HashAlgorithm &H) const {
+   StringRef InputFilename = Config.InputFileNames.front();
+ 
+-  constexpr size_t LargestHeaderSize = CompressedOffloadBundle::V3HeaderSize;
+-
+   ErrorOr<std::unique_ptr<MemoryBuffer>> MaybeInputBuffer =
+-      MemoryBuffer::getFileSlice(InputFilename, LargestHeaderSize, 0);
+      MemoryBuffer::getFile(InputFilename);
+   if (!MaybeInputBuffer) {
+     std::error_code EC = MaybeInputBuffer.getError();
+     return createStringError(EC, Twine("Failed to open ") + InputFilename +
+@@ -167,14 +165,17 @@ Error UnbundleCommand::addInputIdentifier(HashAlgorithm &H) const {
+ 
+   MemoryBuffer &InputBuffer = **MaybeInputBuffer;
+ 
+-  uint8_t Header[LargestHeaderSize];
+-  memset(Header, 0, sizeof(Header));
+-  memcpy(Header, InputBuffer.getBufferStart(),
+-         std::min(LargestHeaderSize, InputBuffer.getBufferSize()));
+-
+-  // only hash the input file, not the whole header. Colissions are unlikely
+-  // since the header includes a hash (weak) of the contents
+-  H.update(Header);
+  using Header = CompressedOffloadBundle::CompressedBundleHeader;
+  Expected<Header> MaybeHeader = Header::tryParse(InputBuffer.getBuffer());
+  if (!MaybeHeader)
+    return MaybeHeader.takeError();
+
+    // The hash represents the contents of the bundle. Extracting the same
+  // contents should give the same result, regardless of the compression
+  // algorithm or header version. Since the hash used by the offload bundler is
+  // not a cryptographic hash, we also add the uncompressed file size.
+  H.update(MaybeHeader->Hash);
+  H.update(MaybeHeader->UncompressedFileSize);
+   return Error::success();
+ }
+ 
+diff --git a/amd/comgr/src/comgr-compiler.cpp b/amd/comgr/src/comgr-compiler.cpp
+index 82102910a..96ca28b94 100644
+--- a/amd/comgr/src/comgr-compiler.cpp
+++ b/amd/comgr/src/comgr-compiler.cpp
+@@ -70,6 +70,7 @@
+ #include "llvm/MC/MCCodeEmitter.h"
+ #include "llvm/MC/MCContext.h"
+ #include "llvm/MC/MCInstrInfo.h"
+#include "llvm/MC/MCInstPrinter.h"
+ #include "llvm/MC/MCObjectFileInfo.h"
+ #include "llvm/MC/MCObjectWriter.h"
+ #include "llvm/MC/MCParser/MCAsmParser.h"
+@@ -462,8 +463,9 @@ bool executeAssemblerImpl(AssemblerInvocation &Opts, DiagnosticsEngine &Diags,
+ 
+   // FIXME: There is a bit of code duplication with addPassesToEmitFile.
+   if (Opts.OutputType == AssemblerInvocation::FT_Asm) {
+-    MCInstPrinter *IP = TheTarget->createMCInstPrinter(
+-        llvm::Triple(Opts.Triple), Opts.OutputAsmVariant, *MAI, *MCII, *MRI);
+    std::unique_ptr<MCInstPrinter> InstructionPrinter(
+      TheTarget->createMCInstPrinter(
+         llvm::Triple(Opts.Triple), Opts.OutputAsmVariant, *MAI, *MCII, *MRI));
+     std::unique_ptr<MCCodeEmitter> MCE;
+     std::unique_ptr<MCAsmBackend> MAB;
+     if (Opts.ShowEncoding) {
+@@ -472,7 +474,7 @@ bool executeAssemblerImpl(AssemblerInvocation &Opts, DiagnosticsEngine &Diags,
+       MAB.reset(TheTarget->createMCAsmBackend(*STI, *MRI, Options));
+     }
+     auto FOut = std::make_unique<formatted_raw_ostream>(*Out);
+-    Str.reset(TheTarget->createAsmStreamer(Ctx, std::move(FOut), IP,
+    Str.reset(TheTarget->createAsmStreamer(Ctx, std::move(FOut), std::move(InstructionPrinter),
+                                            std::move(MCE), std::move(MAB)));
+   } else if (Opts.OutputType == AssemblerInvocation::FT_Null) {
+     Str.reset(createNullStreamer(Ctx));
+@@ -653,9 +655,9 @@ void logArgv(raw_ostream &OS, StringRef ProgramName,
+ amd_comgr_status_t executeCommand(const Command &Job, raw_ostream &LogS,
+                                   DiagnosticOptions &DiagOpts,
+                                   llvm::vfs::FileSystem &FS) {
+-  TextDiagnosticPrinter DiagClient(LogS, &DiagOpts);
+  TextDiagnosticPrinter DiagClient(LogS, DiagOpts);
+   IntrusiveRefCntPtr<DiagnosticIDs> DiagID(new DiagnosticIDs);
+-  DiagnosticsEngine Diags(DiagID, &DiagOpts, &DiagClient, false);
+  DiagnosticsEngine Diags(DiagID, DiagOpts, &DiagClient, false);
+ 
+   auto Arguments = Job.getArguments();
+   SmallVector<const char *, 128> Argv;
+@@ -750,7 +752,7 @@ AMDGPUCompiler::executeInProcessDriver(ArrayRef<const char *> Args) {
+   // here is mostly copy-and-pasted from driver.cpp/cc1_main.cpp/various Clang
+   // tests to try to approximate the same behavior as running the `clang`
+   // executable.
+-  IntrusiveRefCntPtr<DiagnosticOptions> DiagOpts(new DiagnosticOptions);
+  std::unique_ptr<DiagnosticOptions> DiagOpts(new DiagnosticOptions);
+   unsigned MissingArgIndex, MissingArgCount;
+   InputArgList ArgList = getDriverOptTable().ParseArgs(
+       Args.slice(1), MissingArgIndex, MissingArgCount);
+@@ -759,9 +761,9 @@ AMDGPUCompiler::executeInProcessDriver(ArrayRef<const char *> Args) {
+   // DiagnosticsEngine actually exists.
+   (void)ParseDiagnosticArgs(*DiagOpts, ArgList);
+   TextDiagnosticPrinter *DiagClient =
+-      new TextDiagnosticPrinter(LogS, &*DiagOpts);
+      new TextDiagnosticPrinter(LogS, *DiagOpts);
+   IntrusiveRefCntPtr<DiagnosticIDs> DiagID(new DiagnosticIDs);
+-  DiagnosticsEngine Diags(DiagID, &*DiagOpts, DiagClient);
+  DiagnosticsEngine Diags(DiagID, *DiagOpts, DiagClient);
+ 
+   ProcessWarningOptions(Diags, *DiagOpts, *OverlayFS, /*ReportDiags=*/false);
+ 
@@ -0,0 +1,55 @@
+##Fix issue with HIP, where compilation flags are incorrect, see issue:
+#https://github.com/RadeonOpenCompute/ROCm-CompilerSupport/issues/49
+#Remove redundant includes:
+sed -i '/Args.push_back("-isystem");/,+3d' amd/comgr/src/comgr-compiler.cpp
+#Source hard codes the libdir too:
+sed -i 's/lib\(\/clang\)/%{_lib}\1/' amd/comgr/src/comgr-compiler.cpp
+
+# Unsupported options
+sed -i 's@Args.push_back("-mlink-builtin-bitcode-postopt");@//Args.push_back("-mlink-builtin-bitcode-postopt");@' amd/comgr/src/comgr-compiler.cpp
+
+# Use system perl
+sed -i 's|\(/usr/bin/\)env perl|\1perl|' amd/hipcc/bin/hipvars.pm
+
+# Default rocm path is _prefix
+# HIPCC fixes to find clang++
+
+sed -i 's| or -e "$HIP_PATH/bin/clang"||' amd/hipcc/bin/hipvars.pm
+sed -i 's|lib/llvm/bin|%{_lib}/llvm%{llvm_maj_ver}/bin|' \
+    amd/hipcc/bin/hipvars.pm amd/hipcc/src/hipBin_amd.h amd/hipcc/src/hipBin_base.h
+sed -i 's|-e "$HIP_ROCCLR_HOME/bin/clang" or ||' amd/hipcc/bin/hipvars.pm
+
+# Fixup finding /opt/llvm
+sed -i -e 's@sys::path::append(LLVMPath, "llvm");@//sys::path::append(LLVMPath, "llvm");@' amd/comgr/src/comgr-env.cpp
+# Fixup finding /opt/rocm/hip
+sed -i -e 's@sys::path::append(HIPPath, "hip");@//sys::path::append(HIPPath, "hip");@' amd/comgr/src/comgr-env.cpp
+
+# Default rocm path is _prefix
+sed -i -e 's@/opt/rocm@%{_prefix}@' amd/hipcc/src/hipBin_base.h
+
+LLVM_BINDIR=`llvm-config --bindir`
+if [ ! -x ${LLVM_BINDIR}/clang++ ]; then
+    echo "Something wrong with llvm-config"
+    false
+fi
+echo "s@\$ROCM_PATH/lib/llvm/bin@${LLVM_BINDIR}@" > pm.sed
+echo "s@hipClangPath /= \"lib/llvm/bin\"@hipClangPath = \"${LLVM_BINDIR}\"@" > h.sed
+sed -i -f pm.sed amd/hipcc/bin/hipvars.pm
+sed -i -f h.sed amd/hipcc/src/hipBin_amd.h
+
+# ROCm upstream uses /opt for rocm-runtime, but we uses /usr
+# Don't include it again since /usr/include is already included:
+sed -i '/" -isystem " + hsaPath + "\/include"/d' amd/hipcc/src/hipBin_amd.h
+
+sed -i 's/find_package(Clang REQUIRED CONFIG)/find_package(Clang REQUIRED)/' amd/comgr/CMakeLists.txt
+sed -i 's/find_package(LLD REQUIRED CONFIG)/find_package(LLD REQUIRED)/' amd/comgr/CMakeLists.txt
+sed -i 's@${CLANG_CMAKE_DIR}/../../../@/usr/lib/clang/%{llvm_maj_ver}/@' amd/comgr/cmake/opencl_pch.cmake
+
+# CMP0053 OLD is only needed on Windows. But on new version of cmake it is deprecated.
+sed -i 's/cmake_policy(SET CMP0053 OLD)/cmake_policy(SET CMP0053 NEW)/' amd/device-libs/cmake/OCL.cmake
+
+# Fix up the path to the device libs hipcc uses
+sed -i -e 's@amdgcnBitcode = roccmPath@amdgcnBitcode = "%{_prefix}/%{amd_device_libs_prefix}/"@' amd/hipcc/src/hipBin_amd.h
+
+# Fix up the location AMD_DEVICE_LIBS_PREFIX
+sed -i 's|@AMD_DEVICE_LIBS_PREFIX_CODE@|set(AMD_DEVICE_LIBS_PREFIX "%{_prefix}/%{amd_device_libs_prefix}")|' amd/device-libs/AMDDeviceLibsConfig.cmake.in
@@ -0,0 +1,298 @@
+# SPDX-FileCopyrightText: (C) 2025, 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2025, 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+# SPDX-FileContributor: misaka00251 <liuxin@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# The package follows LLVM's major version, but API version is still important:
+%global comgr_maj_api_ver 3
+%global comgr_full_api_ver %{comgr_maj_api_ver}.0
+
+# What LLVM is upstream using (use LLVM_VERSION_MAJOR from cmake/Modules/LLVMVersion.cmake):
+%global llvm_maj_ver 21
+# Sakura286: ROCm 7.1.1 uses LLVM 20, but only LLVM 21 is on openRuyi.
+#            Backport is needed.
+%global rocm_llvm_maj_ver 20
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+%global bundle_prefix %{_libdir}/llvm%{llvm_maj_ver}
+%global llvm_triple %{_target_platform}
+%global amd_device_libs_prefix lib/clang/%{llvm_maj_ver}
+
+%global toolchain clang
+
+%ifarch x86_64
+%global targets_to_build "X86;AMDGPU"
+%endif
+%ifarch riscv64
+%global targets_to_build "RISCV;AMDGPU"
+%endif
+
+# All the tests are not enabled both on fedora and debian
+# https://salsa.debian.org/rocm-team/rocm-llvm/-/blob/debian/unstable/debian/rules
+# https://src.fedoraproject.org/rpms/rocm-compilersupport/blob/rawhide/f/rocm-compilersupport.spec
+# Disabled by default.
+%bcond device_libs_test 0
+%bcond comgr_test 0
+
+Name:           rocm-llvm
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        Various AMD ROCm LLVM related services
+# llvm is Apache-2.0 WITH LLVM-exception OR NCSA
+# hipcc is MIT, comgr and device-libs are NCSA:
+License:        (Apache-2.0 WITH LLVM-exception OR NCSA) AND NCSA AND MIT
+URL:            https://github.com/ROCm/llvm-project
+#!RemoteAsset:  sha256:d76a16db4a56914383029e241823f7bc2a3d645f2967dd22230f11c11cfe189e
+Source0:        %{url}/archive/refs/tags/rocm-%{rocm_version}.tar.gz
+Source1:        rocm-llvm.prep.in
+
+# RISC-V support patches
+# https://salsa.debian.org/rocm-team/rocm-llvm/-/merge_requests/2
+Patch0:         0002-Use-signed-char-in-comgr-building.patch
+# Backport mainline comgr patches since 7.1.1 is build on llvm-20
+Patch1:         0003-adapt-comgr-api-to-llvm-21.patch
+
+BuildRequires:  clang >= %{llvm_maj_ver}
+BuildRequires:  clang-devel >= %{llvm_maj_ver}
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  fdupes
+BuildRequires:  lld >= %{llvm_maj_ver}
+BuildRequires:  lld-devel >= %{llvm_maj_ver}
+BuildRequires:  llvm-devel >= %{llvm_maj_ver}
+BuildRequires:  llvm-test >= %{llvm_maj_ver}
+BuildRequires:  pkgconfig(libffi)
+BuildRequires:  pkgconfig(libxml-2.0)
+BuildRequires:  pkgconfig(libzstd)
+# comgr requires python
+BuildRequires:  pkgconfig(python3)
+BuildRequires:  pkgconfig(zlib)
+BuildRequires:  rocm-cmake >= %{rocm_release}
+
+%description
+%{summary}
+
+%package        macros
+Summary:        ROCm Compiler RPM macros for RPM Build
+BuildArch:      noarch
+
+%description    macros
+This package contains ROCm compiler related RPM macros.
+
+%package     -n rocm-device-libs
+Summary:        AMD ROCm LLVM bit code libraries
+
+%description -n rocm-device-libs
+This package contains a set of AMD specific device-side language runtime
+libraries in the form of bit code. Specifically:
+ - Open Compute library controls
+ - Open Compute Math library
+ - Open Compute Kernel library
+ - OpenCL built-in library
+ - HIP built-in library
+ - Heterogeneous Compute built-in library
+
+%package     -n rocm-comgr
+Summary:        AMD ROCm LLVM Code Object Manager
+Provides:       comgr(major) = %{comgr_maj_api_ver}
+Provides:       rocm-comgr = %{comgr_full_api_ver}-%{release}
+
+%description -n rocm-comgr
+The AMD Code Object Manager (Comgr) is a shared library which provides
+operations for creating and inspecting code objects.
+
+%package     -n rocm-comgr-devel
+Summary:        AMD ROCm LLVM Code Object Manager
+Requires:       rocm-comgr%{?_isa} = %{version}-%{release}
+Requires:       rocm-device-libs
+
+%description -n rocm-comgr-devel
+The AMD Code Object Manager (Comgr) development package.
+
+%package     -n hipcc
+Summary:        HIP compiler driver
+Requires:       rocm-device-libs = %{version}-%{release}
+Suggests:       rocminfo
+
+%description -n hipcc
+hipcc is a compiler driver utility that will call clang or nvcc, depending on
+target, and pass the appropriate include and library options for the target
+compiler and HIP infrastructure.
+
+hipcc will pass-through options to the target compiler. The tools calling hipcc
+must ensure the compiler options are appropriate for the target compiler.
+
+%prep
+%autosetup -p1 -n llvm-project-rocm-%{rocm_version}
+
+# llvm_maj_ver sanity check (we should be matching the bundled llvm major ver):
+if ! grep -q "set(LLVM_VERSION_MAJOR %{llvm_maj_ver})" cmake/Modules/LLVMVersion.cmake; then
+    echo "ERROR llvm_maj_ver macro is not correctly set"
+    # Sakura286: ROCm 7.1.1 uses LLVM 20, but only 21 is on openRuyi. Sad.
+    # TODO: Need to re-enable this 'if' when rocm upstream bump to llvm-21
+    # exit 1
+fi
+
+# Make sure we only build the AMD bits by discarding the bundled llvm code:
+ls | grep -xv "amd" | xargs rm -r
+
+install -pm 755 %{SOURCE1} prep.sh
+sed -i -e 's@%%{_prefix}@%{_prefix}@' prep.sh
+sed -i -e 's@%%{_lib}@%{_lib}@' prep.sh
+sed -i -e 's@%%{amd_device_libs_prefix}@%{amd_device_libs_prefix}@' prep.sh
+sed -i -e 's@%%{bundle_prefix}@%{bundle_prefix}@' prep.sh
+sed -i -e 's@%%{llvm_maj_ver}@%{llvm_maj_ver}@' prep.sh
+grep -v '%%{' prep.sh
+
+. ./prep.sh
+
+%build
+CLANG_VERSION=%llvm_maj_ver
+
+# Maybe use llvm-config-%{llvm_maj_ver} in the future
+LLVM_BINDIR=`%{_libdir}/llvm%{llvm_maj_ver}/bin/llvm-config --bindir`
+LLVM_CMAKEDIR=`%{_libdir}/llvm%{llvm_maj_ver}/bin/llvm-config --cmakedir`
+# Only enable one target to accelerate build
+GPU_TARGET="gfx1100;gfx1101;gfx1200;gfx1201"
+
+echo "%%rocmllvm_version $CLANG_VERSION"        >  macros.rocmcompiler
+echo "%%rocmllvm_bindir $LLVM_BINDIR"           >> macros.rocmcompiler
+echo "%%rocmllvm_cmakedir $LLVM_CMAKEDIR"       >> macros.rocmcompiler
+echo "%%rocm_gpu_list_default \"$GPU_TARGET\""  >> macros.rocmcompiler
+
+export PATH=%{_libdir}/llvm%{llvm_maj_ver}/bin:$PATH
+export INCLUDE_PATH=%{_libdir}/llvm%{llvm_maj_ver}/include
+
+# Build device-libs first, hipcc and comgr need it
+%define _vpath_srcdir amd/device-libs
+%define _vpath_builddir build-devicelibs
+# Workaround for bug in cmake tests not finding amdgcn:
+ln -s %{amd_device_libs_prefix}/amdgcn amdgcn
+#TODO ROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_* should be removed in ROCm 7.0:
+%cmake -DROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_NEW="%{amd_device_libs_prefix}/amdgcn" \
+    -DROCM_DEVICE_LIBS_BITCODE_INSTALL_LOC_OLD="" \
+    -DCMAKE_EXE_LINKER_FLAGS:STRING="-fuse-ld=lld" \
+    %{?__cmake_build_type:-DCMAKE_BUILD_TYPE="%{__cmake_build_type}"}
+%cmake_build -- %{?_smp_mflags}
+# Used by comgr to find device libs when building:
+export ROCM_PATH=$(realpath %__cmake_builddir)
+
+# Build comgr
+%define _vpath_srcdir amd/comgr
+%define _vpath_builddir build-comgr
+%cmake -DCMAKE_PREFIX_PATH=$ROCM_PATH \
+    -DCMAKE_MODULE_PATH=%{_libdir}/llvm%{llvm_maj_ver}/lib \
+    -DCMAKE_BUILD_TYPE="RELEASE" \
+    -DCMAKE_EXE_LINKER_FLAGS:STRING="-fuse-ld=lld" \
+    -DBUILD_TESTING=%{?with_comgr_test:ON}%{!?with_comgr_test:OFF}
+%cmake_build -- %{?_smp_mflags}
+
+# Build hipcc
+%define _vpath_srcdir amd/hipcc
+%define _vpath_builddir build-hipcc
+%cmake -DHIPCC_BACKWARD_COMPATIBILITY=OFF \
+    -DCMAKE_EXE_LINKER_FLAGS:STRING="-fuse-ld=lld"
+%cmake_build -- %{?_smp_mflags}
+
+%check
+# Test device-libs
+%define _vpath_srcdir amd/device-libs
+%define _vpath_builddir build-devicelibs
+# Workaround for bug in cmake tests not finding amdgcn:
+ln -s %{amd_device_libs_prefix}/amdgcn build-devicelibs/amdgcn
+# Below tests are failed:
+#    6 - compile_native_rcp__gfx600 (Failed)
+#    7 - compile_native_rsqrt__gfx600 (Failed)
+#   10 - compile_native_rcp__gfx700 (Failed)
+#   11 - compile_native_rsqrt__gfx700 (Failed)
+#   14 - compile_native_rcp__gfx803 (Failed)
+#   15 - compile_native_rsqrt__gfx803 (Failed)
+#   18 - compile_atomic_work_item_fence__gfx803 (Failed)
+#   19 - compile_atomic_work_item_fence__gfx900 (Failed)
+#   20 - compile_atomic_work_item_fence__gfx90a (Failed)
+#   21 - compile_atomic_work_item_fence__gfx1030 (Failed)
+#   22 - compile_atomic_work_item_fence__gfx1100 (Failed)
+#   23 - compile_atomic_work_item_fence__gfx1200 (Failed)
+%{?with_device_libs_test:%ctest}
+
+# Test comgr
+%define _vpath_srcdir amd/comgr
+%define _vpath_builddir build-comgr
+# Below tests are failed:
+#    2 - comgr_disasm_llvm_reloc_test (SEGFAULT)
+#    3 - comgr_disasm_llvm_so_test (SEGFAULT)
+#    5 - comgr_disasm_options_test (SEGFAULT)
+#   13 - comgr_compile_test (Failed)
+#   14 - comgr_compile_minimal_test (Failed)
+#   16 - comgr_compile_log_remarks_test (Failed)
+#   17 - comgr_compile_source_with_device_libs_to_bc_with_vfs_test (Failed)
+#   21 - comgr_get_data_isa_name_test (Failed)
+#   29 - comgr_mangled_names_test (Failed)
+#   30 - comgr_multithread_test (SEGFAULT)
+#   32 - comgr_compile_hip_test (Failed)
+#   33 - comgr_compile_hip_to_relocatable (Failed)
+#   34 - comgr_mangled_names_hip_test (Failed)
+#   35 - comgr_unbundle_hip_test (Failed)
+%{?with_comgr_test:%ctest}
+
+%install
+# Install macros
+install -Dpm 644 macros.rocmcompiler \
+    %{buildroot}%{_rpmmacrodir}/macros.rocmcompiler
+
+# Install device-libs
+%define _vpath_builddir build-devicelibs
+%cmake_install
+
+# Install comgr
+%define _vpath_builddir build-comgr
+%cmake_install
+
+# Install hipcc
+%define _vpath_builddir build-hipcc
+%cmake_install
+
+rm -f %{buildroot}%{_datadir}/doc/ROCm-Device-Libs/LICENSE.TXT
+rm -rf %{buildroot}%{_datadir}/doc/amd_comgr
+rm -f %{buildroot}%{_datadir}/doc/hipcc/LICENSE.txt
+rm -f %{buildroot}%{_datadir}/doc/hipcc/README.md
+
+%files macros
+%{_rpmmacrodir}/macros.rocmcompiler
+
+%files -n rocm-device-libs
+%doc        amd/device-libs/README.md amd/device-libs/doc/*.md
+%license    amd/device-libs/LICENSE.TXT
+%dir        %{_libdir}/cmake/AMDDeviceLibs
+%{_libdir}/cmake/AMDDeviceLibs/*.cmake
+%{_prefix}/%{amd_device_libs_prefix}/amdgcn
+
+%files -n rocm-comgr
+%doc        amd/comgr/README.md
+%license    amd/comgr/LICENSE.txt
+%license    amd/comgr/NOTICES.txt
+%{_libdir}/libamd_comgr.so.*
+
+%files -n rocm-comgr-devel
+%dir        %{_includedir}/amd_comgr
+%dir        %{_libdir}/cmake/amd_comgr
+%{_includedir}/amd_comgr/amd_comgr.h
+%{_libdir}/libamd_comgr.so
+%{_libdir}/cmake/amd_comgr/*.cmake
+
+%files -n hipcc
+%doc        amd/hipcc/README.md
+%license    amd/hipcc/LICENSE.txt
+%license    amd/hipcc/README.md
+%{_bindir}/hipcc
+%{_bindir}/hipconfig
+%{_bindir}/hipvars.pm
+
+%changelog
+%{?autochangelog}
@@ -0,0 +1,32 @@
+From 3d470b52d1cc06a3e33cbf38bdde6d3f4f008a25 Mon Sep 17 00:00:00 2001
+From: Tom Rix <Tom.Rix@amd.com>
+Date: Mon, 3 Nov 2025 06:33:31 -0800
+Subject: [PATCH] rocm-origami remove scope for variables
+
+---
+ cmake/origami-config.cmake.in | 12 ------------
+ 1 file changed, 12 deletions(-)
+
+diff --git a/cmake/origami-config.cmake.in b/cmake/origami-config.cmake.in
+index d6c8000d0261..19370e7dd5bd 100644
+--- a/cmake/origami-config.cmake.in
+++ b/cmake/origami-config.cmake.in
+@@ -6,15 +6,4 @@ find_dependency(hip REQUIRED)
+ 
+ include("${CMAKE_CURRENT_LIST_DIR}/origami-targets.cmake")
+ 
+-block(SCOPE_FOR VARIABLES)
+-    if(NOT TARGET origami::origami)
+-        message(FATAL_ERROR "origami::origami target is missing")
+-    endif()
+-    
+-    get_target_property(link_libraries origami::origami INTERFACE_LINK_LIBRARIES)
+-
+-    if(link_libraries AND "hip::device" IN_LIST link_libraries)
+-        message(FATAL_ERROR "Do not export targets with hip::device as an interface link library")
+-    endif()
+-endblock()
+ 
+-- 
+2.51.0
+
@@ -0,0 +1,81 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+Name:           rocm-origami
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        Analytical GEMM Solution Selection
+License:        MIT
+Url:            https://github.com/ROCm/rocm-libraries
+#!RemoteAsset:  sha256:1fb56e620a06e198aeec2cf37c11e6879d0c67c62e295b48779b7f486e34acb4
+Source0:        %{url}/releases/download/rocm-%{version}/origami.tar.gz
+# License file is not included in the release tarball
+#!RemoteAsset:  sha256:b185aaa652b0bf066c37a0d6314ce4bf4521e4a3c9bf46edd2f6a777ac522223
+Source1:        https://raw.githubusercontent.com/ROCm/rocm-libraries/develop/shared/origami/LICENSE.md
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON
+
+# Workaround hipblaslt build issue:
+#   origami::origami target is missing
+# https://github.com/ROCm/rocm-libraries/issues/2422
+Patch0:         0001-rocm-origami-remove-scope-for-variables.patch
+
+BuildRequires:  clang
+BuildRequires:  cmake
+BuildRequires:  cmake(hip)
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-llvm-macros
+BuildRequires:  ninja
+
+%description
+The name "origami" still evokes the elegance of transforming
+a flat (2-D) sheet into intricate higher dimensional
+structures. In this context, however, Origami has evolved
+into a tool set for GEMM solution selection and optimization.
+Inspired by the art of paper folding, the library now enables
+users to explore a range of tiling and mapping configurations
+and to make informed decisions on data and computation mapping
+for high-performance GEMM operations.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%prep -a
+# License file is not in the tarball
+cp %{SOURCE1} .
+# Use system rocm-cmake, no downloading
+sed -i -e 's@if(NOT ROCM_FOUND)@if(FALSE)@' cmake/dependencies.cmake
+# We are building from a tarball, not a git repo
+sed -i -e 's@find_package(Git REQUIRED)@#find_package(Git REQUIRED)@' cmake/dependencies.cmake
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/origami/LICENSE.md
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/liborigami.so.0{,.*}
+
+%files devel
+%{_includedir}/origami/
+%{_libdir}/cmake/origami/
+%{_libdir}/liborigami.so
+
+%changelog
+%autochangelog
@@ -1,69 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global rocm_version 7.1.1
-
-Name:           rocprim
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        ROCm parallel primatives
-License:        MIT AND BSD-3-Clause
-URL:            https://github.com/ROCm/rocPRIM
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-BuildOption(conf):  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
-BuildOption(conf):  -DBUILD_TEST=OFF
-BuildOption(conf):  -DCMAKE_AR=%{rocmllvm_bindir}/llvm-ar
-BuildOption(conf):  -DCMAKE_C_COMPILER=%{rocmllvm_bindir}/clang
-BuildOption(conf):  -DCMAKE_CXX_COMPILER=%{rocmllvm_bindir}/clang++
-BuildOption(conf):  -DCMAKE_LINKER=%{rocmllvm_bindir}/ld.lld
-BuildOption(conf):  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/..
-BuildOption(conf):  -DCMAKE_RANLIB=%{rocmllvm_bindir}/llvm-ranlib
-BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
-BuildOption(conf):  -DROCM_SYMLINK_LIBS=OFF
-
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  cmake
-BuildRequires:  cmake(amd_comgr)
-BuildRequires:  cmake(Clang)
-BuildRequires:  cmake(hip)
-BuildRequires:  cmake(hsa-runtime64)
-BuildRequires:  cmake(LLD)
-BuildRequires:  cmake(LLVM)
-BuildRequires:  gcc-c++
-BuildRequires:  python3
-BuildRequires:  rocm-cmake
-BuildRequires:  rocm-device-libs
-BuildRequires:  rocm-llvm-macros
-BuildRequires:  rocminfo
-
-%description
-The rocPRIM is a header-only library providing HIP parallel primitives
-for developing performant GPU-accelerated code on AMD ROCm platform.
-
-%package        devel
-Summary:        ROCm parallel primatives
-BuildArch:      noarch
-
-%description    devel
-The rocPRIM is a header-only library providing HIP parallel primitives
-for developing performant GPU-accelerated code on AMD ROCm platform.
-
-%install -a
-rm -f %{buildroot}%{_prefix}/share/doc/rocprim/LICENSE.md
-
-%files devel
-%doc README.md
-%license LICENSE.md
-%license NOTICES.txt
-%{_includedir}/%{name}
-%{_libdir}/cmake/rocprim
-
-%changelog
-%{?autochangelog}
@@ -1,42 +0,0 @@
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 2fdf5c4..0aac139 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -78,7 +78,8 @@ set(CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake ${PROJECT_SOURCE_DIR}/cmake/Mo
- include(GNUInstallDirs) # install directories
- 
- # ROCm does not use lib64
-set(CMAKE_INSTALL_LIBDIR "lib")
-+# But distributions use
-+# set(CMAKE_INSTALL_LIBDIR "lib")
- 
- include(rocprofiler_register_utilities) # various functions/macros
- include(rocprofiler_register_interfaces) # interface libraries
-@@ -113,6 +114,7 @@ if(ROCPROFILER_REGISTER_BUILD_SAMPLES)
-     add_subdirectory(samples)
- endif()
- 
-include(rocprofiler_register_config_packaging)
-+# packaging don't need cpack
-+# include(rocprofiler_register_config_packaging)
- 
- rocprofiler_register_print_features()
-diff --git a/source/lib/rocprofiler-register/CMakeLists.txt b/source/lib/rocprofiler-register/CMakeLists.txt
-index e15fa88..6d9591e 100644
--- a/source/lib/rocprofiler-register/CMakeLists.txt
-+++ b/source/lib/rocprofiler-register/CMakeLists.txt
-@@ -20,10 +20,13 @@ target_include_directories(
- target_link_libraries(
-     rocprofiler-register
-     PUBLIC rocprofiler-register::headers
-    PRIVATE fmt::fmt glog::glog rocprofiler-register::build-flags
-+    PRIVATE fmt glog rocprofiler-register::build-flags
-             rocprofiler-register::memcheck rocprofiler-register::stdcxxfs
-             rocprofiler-register::dl)
- 
-+target_compile_definitions(rocprofiler-register
-+    PRIVATE GLOG_USE_GLOG_EXPORT)
-+
- set_target_properties(
-     rocprofiler-register
-     PROPERTIES OUTPUT_NAME rocprofiler-register
@@ -1,78 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global rocm_version 7.1.0
-
-Name:           rocprofiler-register
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        A rocprofiler helper library
-License:        MIT AND BSD-3-Clause
-Url:            https://github.com/ROCm/rocprofiler-register
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-# Use system glog
-Patch0:         0001-use-system-buildreq.patch
-
-BuildOption(conf):  -DROCPROFILER_REGISTER_BUILD_FMT=OFF
-BuildOption(conf):  -DROCPROFILER_REGISTER_BUILD_GLOG=OFF
-
-BuildRequires:  cmake
-BuildRequires:  cmake(glog)
-BuildRequires:  gcc-c++
-BuildRequires:  pkgconfig(fmt)
-BuildRequires:  pkgconfig(gflags)
-
-%description
-The rocprofiler-register library is a helper library that coordinates
-the modification of the intercept API table(s) of the HSA/HIP/ROCTx
-runtime libraries by the ROCprofiler (v2) library. The purpose of this
-library is to provide a consistent and automated mechanism of enabling
-performance analysis in the ROCm runtimes which does not rely on
-environment variables or unique methods for each runtime library.
-
-When a runtime is initialized (either explicitly and lazily) and the
-intercept API table is constructed, it passes this API table to
-rocprofiler-register. Rocprofiler-register scans the symbols in the
-address space and if it detects there is at least one visible symbol
-named rocprofiler_configure (which is a function provided by tools),
-it passes the intercept API table to the rocprofiler library (dlopening
-the rocprofiler library if it is not already loaded). The rocprofiler
-library then does an extensive scan for all the instances of the
-rocprofiler_configure symbols and invokes each of them. The
-rocprofiler_configure function (again, provided by a tool) returns
-effectively tells rocprofiler which behaviors it wants to be notified
-about, features it wants to use (e.g. API tracing, kernel dispatch
-timing), etc.
-
-%package        devel
-Summary:        The development package for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    devel
-%{summary}
-
-%install -a
-# Do not install the test source etc
-rm -rf %{buildroot}%{_prefix}/share/rocprofiler-register
-rm -rf %{buildroot}%{_prefix}/share/modulefiles
-rm -rf %{buildroot}%{_prefix}/share/doc/rocprofiler-register/LICENSE.md
-
-%files
-%license LICENSE.md
-%{_libdir}/librocprofiler-register.so.0{,.*}
-
-%files devel
-%doc README.md
-%{_includedir}/rocprofiler-register/
-%{_libdir}/librocprofiler-register.so
-%{_libdir}/cmake/rocprofiler-register/
-
-%changelog
-%{?autochangelog}
@@ -0,0 +1,98 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: Sakura286 <chenxuan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# rocRAND need a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU, so make it optional.
+%bcond test 0
+%if %{with test}
+%global build_test ON
+%else
+%global build_test OFF
+%endif
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm builds with clang
+%global toolchain clang
+
+Name:           rocrand
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm random number generator
+License:        MIT AND BSD-3-Clause
+Url:            https://github.com/ROCm/rocRAND
+#!RemoteAsset:  sha256:15c33c595aa8e4de1d8b3736df9eaf2ceba7914ffebe718f0997b0da28215d9e
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=%{build_test}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+%if %{with test}
+BuildRequires:  cmake(GTest)
+%endif
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+The rocRAND project provides functions that generate pseudo-random and
+quasi-random numbers.
+
+The rocRAND library is implemented in the HIP programming language and
+optimized for AMD's latest discrete GPUs. It is designed to run on top of AMD's
+Radeon Open Compute ROCm runtime, but it also works on CUDA enabled GPUs.
+
+%package        devel
+Summary:        The rocRAND development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+The rocRAND development package.
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/rocrand/LICENSE.md
+
+%files
+%doc README.md
+%license LICENSE.md
+%{_libdir}/librocrand.so.1{,.*}
+
+%files devel
+%{_includedir}/rocrand/
+%{_libdir}/cmake/rocrand/
+%{_libdir}/librocrand.so
+
+%if %{with test}
+%files test
+%{_bindir}/rocRAND/
+%{_bindir}/test_*
+%endif
+
+%changelog
+%autochangelog
@@ -1,49 +0,0 @@
-From 22d2be00dc2289037144abeffab7a5526a8014ea Mon Sep 17 00:00:00 2001
-From: Tom Rix <Tom.Rix@amd.com>
-Date: Thu, 30 Oct 2025 11:27:03 -0700
-Subject: [PATCH] rocsolver ninja job pools
-
---
- CMakeLists.txt | 26 ++++++++++++++++++++++++++
- 1 file changed, 26 insertions(+)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 32757570f70f..003b37f98fc5 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -125,6 +125,32 @@ option(BUILD_SHARED_LIBS "Build rocSOLVER as a shared library" ON)
- include(util)
- include(CheckLanguage)
- include(CMakeDependentOption)
-+#
-+# Seperate linking jobs from compiling
-+# Too many concurrent linking jobs can break the build
-+# Copied from LLVM
-+set(ROCSOLVER_PARALLEL_LINK_JOBS "" CACHE STRING
-+  "Define the maximum number of concurrent link jobs (Ninja only).")
-+if(CMAKE_GENERATOR MATCHES "Ninja")
-+  if(ROCSOLVER_PARALLEL_LINK_JOBS)
-+    set_property(GLOBAL APPEND PROPERTY JOB_POOLS link_job_pool=${ROCSOLVER_PARALLEL_LINK_JOBS})
-+    set(CMAKE_JOB_POOL_LINK link_job_pool)
-+  endif()
-+elseif(ROCSOLVER_PARALLEL_LINK_JOBS)
-+  message(WARNING "Job pooling is only available with Ninja generators.")
-+endif()
-+# Similar for compiling
-+set(ROCSOLVER_PARALLEL_COMPILE_JOBS "" CACHE STRING
-+  "Define the maximum number of concurrent compile jobs (Ninja only).")
-+if(CMAKE_GENERATOR MATCHES "Ninja")
-+  if(ROCSOLVER_PARALLEL_COMPILE_JOBS)
-+    set_property(GLOBAL APPEND PROPERTY JOB_POOLS compile_job_pool=${ROCSOLVER_PARALLEL_COMPILE_JOBS})
-+    set(CMAKE_JOB_POOL_COMPILE compile_job_pool)
-+  endif()
-+elseif(ROCSOLVER_PARALLEL_COMPILE_JOBS)
-+  message(WARNING "Job pooling is only available with Ninja generators.")
-+endif()
-+
- include(CheckCXXCompilerFlag)
- 
- option(BUILD_TESTING "Build rocSOLVER tests" OFF)
-- 
-2.51.0
-
@@ -1,42 +0,0 @@
-From 10affbe2ed6ad66b8a3940bc077161d71d8a8d54 Mon Sep 17 00:00:00 2001
-From: Tom Rix <Tom.Rix@amd.com>
-Date: Thu, 30 Oct 2025 11:38:27 -0700
-Subject: [PATCH] rocsolver parallel jobs
-
---
- CMakeLists.txt             | 3 +++
- library/src/CMakeLists.txt | 4 ++++
- 2 files changed, 7 insertions(+)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 003b37f98fc5..1f93a519e537 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -177,6 +177,9 @@ option(WERROR "Treat warnings as errors" OFF)
- option(BUILD_COMPRESSED_DBG "Enable compressed debug symbols" ON)
- check_cxx_compiler_flag("--offload-compress" CXX_COMPILER_SUPPORTS_OFFLOAD_COMPRESS)
- cmake_dependent_option(BUILD_OFFLOAD_COMPRESS "Build with offload compression" ON CXX_COMPILER_SUPPORTS_OFFLOAD_COMPRESS OFF)
-+check_cxx_compiler_flag("-parallel-jobs=4" CXX_COMPILER_SUPPORTS_PARALLEL_HIP_JOBS)
-+cmake_dependent_option(BUILD_PARALLEL_HIP_JOBS "Build with parallel hip jobs" ON CXX_COMPILER_SUPPORTS_PARALLEL_HIP_JOBS OFF)
-+
- 
- message(STATUS "Tests: ${BUILD_CLIENTS_TESTS}")
- message(STATUS "Benchmarks: ${BUILD_CLIENTS_BENCHMARKS}")
-diff --git a/library/src/CMakeLists.txt b/library/src/CMakeLists.txt
-index b39646ee0f1d..7c5cc98b19ba 100755
--- a/library/src/CMakeLists.txt
-+++ b/library/src/CMakeLists.txt
-@@ -448,6 +448,10 @@ if(BUILD_OFFLOAD_COMPRESS)
-   target_compile_options(rocsolver PRIVATE "--offload-compress")
- endif()
- 
-+if(BUILD_PARALLEL_HIP_JOBS)
-+  target_compile_options(rocsolver PRIVATE "-parallel-jobs=4")
-+endif()
-+
- target_include_directories(rocsolver
-   PUBLIC
-     $<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/library/include>
-- 
-2.51.0
-
@@ -1,109 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%global rocm_version 7.1.1
-
-# consume too much time
-%bcond test 0
-%bcond sample 0
-%bcond benchmark 0
-
-Name:           rocsolver
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        Next generation LAPACK implementation for ROCm platform
-License:        BSD-3-Clause AND BSD-2-Clause
-Url:            https://github.com/ROCm/rocSOLVER
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-BuildOption(conf):  -G Ninja
-BuildOption(conf):  -DCMAKE_CXX_COMPILER=hipcc
-BuildOption(conf):  -DCMAKE_C_COMPILER=clang
-BuildOption(conf):  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar
-BuildOption(conf):  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib
-BuildOption(conf):  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/..
-BuildOption(conf):  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
-BuildOption(conf):  -DROCM_SYMLINK_LIBS=OFF
-BuildOption(conf):  -DHIP_PLATFORM=amd
-BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
-BuildOption(conf):  -DBUILD_OFFLOAD_COMPRESS=ON
-BuildOption(conf):  -DBUILD_PARALLEL_HIP_JOBS=ON
-BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{?with_test:ON}%{!?with_test:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=%{?with_benchmark:ON}%{!?with_benchmark:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_SAMPLES=%{?with_sample:ON}%{!?with_sample:OFF}
-
-# https://github.com/ROCm/rocSOLVER/pull/652
-Patch0:         0001-rocsolver-ninja-job-pools.patch
-# https://github.com/ROCm/rocSOLVER/pull/962
-Patch1:         0001-rocsolver-parallel-jobs.patch
-
-BuildRequires:  clang
-BuildRequires:  clang-tools-extra
-BuildRequires:  cmake
-BuildRequires:  cmake(amd_comgr)
-BuildRequires:  cmake(fmt)
-BuildRequires:  cmake(hip)
-BuildRequires:  cmake(hsa-runtime64)
-BuildRequires:  cmake(rocblas)
-BuildRequires:  cmake(rocprim)
-BuildRequires:  compiler-rt
-BuildRequires:  gcc-c++
-BuildRequires:  hipcc
-BuildRequires:  lld
-BuildRequires:  llvm
-BuildRequires:  ninja
-BuildRequires:  pkgconfig(libzstd)
-BuildRequires:  rocm-cmake
-BuildRequires:  rocm-llvm-macros
-BuildRequires:  rocsparse-devel
-BuildRequires:  rocminfo
-
-Provides:       rocsolver = %{version}-%{release}
-
-%description
-rocSOLVER is a work-in-progress implementation of a subset
-of LAPACK functionality on the ROCm platform.
-
-%package        devel
-Summary:        Libraries and headers for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    devel
-%{summary}
-
-%if %{with test}
-%package        test
-Summary:        Tests for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    test
-%{summary}
-%endif
-
-%install -a
-rm -f %{buildroot}%{_prefix}/share/doc/rocsolver/LICENSE.md
-
-%files
-%license LICENSE.md
-%doc README.md
-%{_libdir}/librocsolver.so.0{,.*}
-
-%files devel
-%{_includedir}/rocsolver/
-%{_libdir}/librocsolver.so
-%{_libdir}/cmake/rocsolver/
-
-%if %{with test}
-%files test
-%{_datadir}/rocsolver/
-%{_bindir}/rocsolver*
-%endif
-
-%changelog
-%{?autochangelog}
@@ -1,134 +0,0 @@
-# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
-# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
-# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
-# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
-#
-# SPDX-License-Identifier: MulanPSL-2.0
-
-%bcond test 0
-
-%global rocm_version 7.1.1
-
-Name:           rocsparse
-Version:        %{rocm_version}
-Release:        %autorelease
-Summary:        SPARSE implementation for ROCm
-License:        MIT
-Url:            https://github.com/ROCm/rocSPARSE
-#!RemoteAsset
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz
-BuildSystem:    cmake
-
-BuildOption(conf):  -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF
-BuildOption(conf):  -DBUILD_WITH_OFFLOAD_COMPRESS=ON
-BuildOption(conf):  -DCMAKE_CXX_COMPILER=hipcc
-BuildOption(conf):  -DCMAKE_C_COMPILER=clang
-BuildOption(conf):  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld
-BuildOption(conf):  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar
-BuildOption(conf):  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib
-BuildOption(conf):  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/..
-BuildOption(conf):  -DHIP_PLATFORM=amd
-BuildOption(conf):  -DROCM_SYMLINK_LIBS=OFF
-BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=%{?with_test:ON}%{!?with_test:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{?with_test:ON}%{!?with_test:OFF}
-BuildOption(conf):  -DBUILD_CLIENTS_TESTS_OPENMP=OFF
-BuildOption(conf):  -DBUILD_FORTRAN_CLIENTS=OFF
-BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
-BuildOption(conf):  -G Ninja
-%if %{with test}
-BuildOption(conf):  -DCMAKE_MATRICES_DIR=%{_builddir}/rocsparse-test-matrices/
-%endif
-
-BuildRequires:  clang
-BuildRequires:  clang-tools-extra
-BuildRequires:  compiler-rt
-BuildRequires:  cmake
-BuildRequires:  cmake(amd_comgr)
-BuildRequires:  cmake(hip)
-BuildRequires:  cmake(hsa-runtime64)
-BuildRequires:  cmake(rocprim)
-BuildRequires:  hipcc
-BuildRequires:  lld
-BuildRequires:  llvm
-BuildRequires:  gcc-c++
-BuildRequires:  ninja
-BuildRequires:  pkgconfig(libzstd)
-BuildRequires:  python3
-BuildRequires:  rocm-cmake
-BuildRequires:  rocm-llvm-macros
-BuildRequires:  rocminfo
-%if %{with test}
-BuildRequires:  cmake(GTest)
-BuildRequires:  cmake(rocblas)
-BuildRequires:  cmake(openmp)
-BuildRequires:  gcc-gfortran
-BuildRequires:  python3dist(pyyaml)
-%endif
-
-Provides:       %{name} = %{version}-%{release}
-
-%description
-rocSPARSE exposes a common interface that provides Basic
-Linear Algebra Subroutines for sparse computation
-implemented on top of AMD's Radeon Open eCosystem Platform
-ROCm runtime and toolchains. rocSPARSE is created using
-the HIP programming language and optimized for AMD's
-latest discrete GPUs.
-
-%package        devel
-Summary:        Libraries and headers for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    devel
-%{summary}
-
-%if %{with test}
-%package        test
-Summary:        Tests for %{name}
-Requires:       %{name}%{?_isa} = %{version}-%{release}
-
-%description    test
-%{summary}
-%endif
-
-%prep -a
-# /usr/include/gtest/internal/gtest-port.h:273:2: error: C++ versions less than C++17 are not supported.
-# Convert the c++14 to c++17
-sed -i -e 's@set(CMAKE_CXX_STANDARD 14)@set(CMAKE_CXX_STANDARD 17)@' {,clients/}CMakeLists.txt
-
-
-%install -a
-rm -f %{buildroot}%{_prefix}/share/doc/rocsparse/LICENSE.md
-
-%if %{with test}
-mkdir -p %{buildroot}/%{_datadir}/rocsparse/matrices
-install -pm 644 %{_builddir}/rocsparse-test-matrices/* %{buildroot}/%{_datadir}/rocsparse/matrices
-%endif
-
-%check
-%if %{with test}
-export LD_LIBRARY_PATH=%{_vpath_builddir}/library:$LD_LIBRARY_PATH
-%{_vpath_builddir}/clients/staging/rocsparse-test
-%endif
-
-%files
-%doc README.md
-%license LICENSE.md
-%{_libdir}/librocsparse.so.1{,.*}
-
-%files devel
-%{_includedir}/rocsparse/
-%{_libdir}/librocsparse.so
-%{_libdir}/cmake/rocsparse/
-
-%if %{with test}
-%files test
-%{_bindir}/rocsparse*
-%{_datadir}/rocsparse/test/rocsparse_*
-%{_datadir}/rocsparse/
-%{_libdir}/rocsparse/
-%{_libexecdir}/rocsparse/
-%endif
-
-%changelog
-%{?autochangelog}
@@ -0,0 +1,108 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# rocThrust needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU, so make it optional.
+%bcond test 0
+%if %{with test}
+%global build_test ON
+%else
+%global build_test OFF
+%endif
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm builds with clang
+%global toolchain clang
+
+Name:           rocthrust
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm Thrust library
+
+Url:            https://github.com/ROCm/rocThrust
+VCS:            git:https://github.com/ROCm/rocThrust.git
+License:        Apache-2.0 AND BSD-2-Clause AND BSD-3-Clause AND BSL-1.0 AND MIT
+# All files are Apache 2.0 with some exceptions:
+# ./cmake contains only files under MIT
+# ./internal/benchmark/*.py are dual licensed Apache 2.0 and Boost 1.0
+# ./thrust/ contain some header files that are Boost 1.0 licensed
+# ./thrust/ contain some headers that are dual Apache 2.0 and Boost 1.0
+# ./thrust/cmake/FindTBB.cmake is public domain
+# ./thrust/detail/allocator/allocator_traits.h is dual Apache 2.0 and MIT
+# ./thrust/detail/complex contains BSD 2 clause licensed headers
+#!RemoteAsset:  sha256:995f9498402f207d04aac1edeb845abea295f6f132151ae1e04a6f0d0dc5edf5
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=%{build_test}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+%if %{with test}
+BuildRequires:  cmake(GTest)
+%endif
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(rocprim)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+Thrust is a parallel algorithm library. This library has been
+ported to HIP/ROCm platform, which uses the rocPRIM library.
+
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+%{summary}
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep -a
+# ROCMExportTargetsHeaderOnly.cmake hardcodes 'lib' as the library directory.
+# Change it to the correct platform-specific library directory.
+sed -i -e 's/ROCM_INSTALL_LIBDIR lib/ROCM_INSTALL_LIBDIR %{_lib}/' cmake/ROCMExportTargetsHeaderOnly.cmake
+
+%install -a
+rm -f %{buildroot}%{_docdir}/rocthrust/LICENSE
+
+%files
+%doc README.md
+%license LICENSE
+%license NOTICES.txt
+
+%files devel
+%{_includedir}/thrust/
+%{_libdir}/cmake/rocthrust/
+
+%if %{with test}
+%files test
+%{_bindir}/test_*
+%{_bindir}/rocthrust/
+%endif
+
+%changelog
+%autochangelog
@@ -0,0 +1,106 @@
+# SPDX-FileCopyrightText: (C) 2026 Institute of Software, Chinese Academy of Sciences (ISCAS)
+# SPDX-FileCopyrightText: (C) 2026 openRuyi Project Contributors
+# SPDX-FileContributor: CHEN Xuan <chenxuan@iscas.ac.cn>
+# SPDX-FileContributor: Yifan Xu <xuyifan@iscas.ac.cn>
+#
+# SPDX-License-Identifier: MulanPSL-2.0
+
+# roctracer needs a GPU to run tests, but we could still
+# keep the test cases for packagers who have a GPU, so make it optional.
+%bcond test 0
+%if %{with test}
+%global build_test ON
+%else
+%global build_test OFF
+%endif
+
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}
+
+# rocm stack builds with clang
+%global toolchain clang
+
+Name:           roctracer
+Version:        %{rocm_version}
+Release:        %autorelease
+Summary:        ROCm Tracer Callback/Activity Library
+Url:            https://github.com/ROCm/roctracer
+VCS:            git:https://github.com/ROCm/roctracer.git
+License:        MIT
+#!RemoteAsset:  sha256:dec80803c6d2d684759172145177849efda65672645b95a2f2ad1a84335043bb
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake
+
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  cmake
+BuildRequires:  cmake(amd_comgr)
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
+BuildRequires:  ninja
+BuildRequires:  pkgconfig(atomic_ops)
+BuildRequires:  python3dist(cppheaderparser)
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros
+
+%description
+roctracer is a callback and activity tracing library for ROCm. It provides
+function call tracing for HIP and other ROCm runtimes, activity (asynchronous)
+tracing, and ROCTx user-defined event markers.
+
+%package        devel
+Summary:        The roctracer development package
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    devel
+The roctracer development package.
+
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
+%endif
+
+%prep -a
+# No knob in cmake to turn off testing
+%if %{without test}
+sed -i -e 's@add_subdirectory(test)@#add_subdirectory(test)@' CMakeLists.txt
+%else
+# Adjust test running script lib dir
+sed -i -e 's@../lib/@../%{_lib}/@' test/run.sh
+%endif
+
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/%{name}/LICENSE.md
+rm -rf %{buildroot}%{_datadir}/doc/%{name}-asan
+
+%files
+%license LICENSE.md
+%doc README.md
+%{_libdir}/libroctracer64.so.*
+%{_libdir}/libroctx64.so.*
+%{_libdir}/roctracer/
+
+%files devel
+%{_includedir}/roctracer/
+%{_libdir}/libroctracer64.so
+%{_libdir}/libroctx64.so
+
+%if %{with test}
+%files test
+%{_datadir}/roctracer/
+%endif
+
+%changelog
+%autochangelog
				`@@ -0,0 +1 @@`
				`u ollama - "Runs Ollama" /var/lib/ollama /sbin/nologin`