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
80 changed files with 4274 additions and 5195 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
@@ -1,21 +1,21 @@
-# For testing
-# Depends on downloading and being in a git repo
-%bcond test 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

-# Header only package
-%global debug_package %{nil}
-
-Summary:        Functional Programming Library for C++
 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
-Version:        0.2.25
-Release:        1%{?dist}
-
-URL:            https://github.com/Dobiasd/FunctionalPlus
-Source0:        %{url}/archive/v%{version}.tar.gz#/FunctionalPlus-%{version}.tar.gz
+#!RemoteAsset:  sha256:8864a3e9bebde6ebed71b49ac2a036cedf9ae0f02ce758bc28c21e6a2ae15803
+Source0:         %{url}/archive/v%{version}.tar.gz
+BuildSystem:    cmake

 BuildRequires:  cmake
-BuildRequires:  gcc-c++

 %description
 FunctionalPlus is a small header-only library supporting you in
@@ -26,51 +26,11 @@ 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.

-%package devel
-
-Summary:        Functional Programming Library for C++
-Provides:       %{name}-static = %{version}-%{release}
-
-%description devel
-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.
-
-%prep
-%autosetup -p1 -n FunctionalPlus-%{version}
-
-# License check flags this as BSD 3-Clause
-# api_search not distributed, remove to make license simpler
-rm -rf api_search
-
-%build
-%cmake 
-%cmake_build
-
-%if %{with test}
-%check
-%ctest
-%endif
-
-%install
-%cmake_install
-
-%files devel
-%dir %_includedir/%{name}
-%dir %_includedir/%{name}/internal
-%dir %_includedir/%{name}/internal/asserts
-%license LICENSE
+%files
 %doc README.md
-%_includedir/%{name}/*_defines
-%_includedir/%{name}/*.hpp
-%_includedir/%{name}/internal/*.hpp
-%_includedir/%{name}/internal/asserts/*.hpp
-%_libdir/cmake/FunctionalPlus/
+%license LICENSE
+%{_includedir}/fplus/
+%{_libdir}/cmake/FunctionalPlus/

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 0.2.25-1
- Import from upstream
+%autochangelog
@@ -1,117 +0,0 @@
-%bcond test 0
-
-# Header only package
-%global debug_package %{nil}
-
-Summary:        Header-only library for using Keras (TensorFlow) models in C++
-Name:           frugally-deep
-License:        MIT
-# Main license is MIT
-# BSD-2-Clause is only for cmake/HunterGate.cmake and that is not distributed
-Version:        0.15.30
-Release:        1%{?dist}
-
-URL:            https://github.com/Dobiasd/frugally-deep
-Source0:        %{url}/archive/v%{version}.tar.gz#/%{name}-%{version}.tar.gz
-
-BuildRequires:  cmake
-BuildRequires:  eigen3
-BuildRequires:  fplus-devel
-#BuildRequires:  nlohmann_json-devel
-BuildRequires:  nlohmann-json
-BuildRequires:  gcc-c++
-
-%description
-Would you like to build/train a model using Keras/Python? And would
-you like to run the prediction (forward pass) on your model in C++
-without linking your application against TensorFlow? Then
-frugally-deep is exactly for you.
-
-frugally-deep
-
-* is a small header-only library written in modern and pure C++.
-* is very easy to integrate and use.
-* depends only on FunctionalPlus, Eigen and json - also header-only
-  libraries.
-* supports inference (model.predict) not only for sequential models
-  but also for computational graphs with a more complex topology,
-  created with the functional API.
-* re-implements a (small) subset of TensorFlow, i.e., the operations
-  needed to support prediction.
-* results in a much smaller binary size than linking against TensorFlow.
-* works out-of-the-box also when compiled into a 32-bit executable.
-  (Of course, 64 bit is fine too.)
-* avoids temporarily allocating (potentially large chunks of)
-  additional RAM during convolutions (by not materializing the im2col
-  input matrix).
-* utterly ignores even the most powerful GPU in your system and uses
-  only one CPU core per prediction. ;-)
-* but is quite fast on one CPU core, and you can run multiple
-  predictions in parallel, thus utilizing as many CPUs as you like
-  to improve the overall prediction throughput of your
-  application/pipeline.
-
-%package devel
-
-Summary:        Header-only library for using Keras (TensorFlow) models in C++
-Provides:       %{name}-static = %{version}-%{release}
-
-%description devel
-Would you like to build/train a model using Keras/Python? And would
-you like to run the prediction (forward pass) on your model in C++
-without linking your application against TensorFlow? Then
-frugally-deep is exactly for you.
-
-frugally-deep
-
-* is a small header-only library written in modern and pure C++.
-* is very easy to integrate and use.
-* depends only on FunctionalPlus, Eigen and json - also header-only
-  libraries.
-* supports inference (model.predict) not only for sequential models
-  but also for computational graphs with a more complex topology,
-  created with the functional API.
-* re-implements a (small) subset of TensorFlow, i.e., the operations
-  needed to support prediction.
-* results in a much smaller binary size than linking against TensorFlow.
-* works out-of-the-box also when compiled into a 32-bit executable.
-  (Of course, 64 bit is fine too.)
-* avoids temporarily allocating (potentially large chunks of)
-  additional RAM during convolutions (by not materializing the im2col
-  input matrix).
-* utterly ignores even the most powerful GPU in your system and uses
-  only one CPU core per prediction. ;-)
-* but is quite fast on one CPU core, and you can run multiple
-  predictions in parallel, thus utilizing as many CPUs as you like
-  to improve the overall prediction throughput of your
-  application/pipeline.
-
-%prep
-%autosetup -p1 -n %{name}-%{version}
-
-# cmake changed
-sed -i -e 's@cmake_minimum_required(VERSION 3.2)@cmake_minimum_required(VERSION 3.5)@' CMakeLists.txt
-
-%build
-%cmake 
-%cmake_build
-
-%if %{with test}
-%check
-%ctest
-%endif
-
-%install
-%cmake_install
-
-%files devel
-%dir %_includedir/fdeep
-%dir %_libdir/cmake/%{name}
-%license LICENSE
-%doc README.md
-%_includedir/fdeep/*
-%_libdir/cmake/%{name}/*
-
-%changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 0.15.30-1
- Import from upstream
@@ -1,64 +1,40 @@
-# there is no debug package
-%global debug_package %{nil}
+# 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:        2.2.0
-Release:        1%{?dist}
-Summary:        A C++ half-precision floating point type
+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

-URL:            http://sourceforge.net/projects/half
-Source0:        %{url}/files/%{name}/%{version}/%{name}-%{version}.zip
-BuildArch:      noarch
-
-BuildRequires:  unzip
+BuildRequires:  cmake
+BuildRequires:  rocm-cmake

 %description
-This is a C++ header-only library to provide an IEEE-754 conformant
-half-precision floating point type along with corresponding arithmetic
-operators, type conversions and common mathematical functions. It aims
-for both efficiency and ease of use, trying to accurately mimic the
-behaviour of the builtin floating point types at the best performance
-possible. It automatically uses and provides C++11 features when
-possible, but stays completely C++98-compatible when neccessary.
+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.

-%package devel
-Summary:        A C++ half-precision floating point type
-Provides:       %{name}-static = %{version}-%{release}
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/half/LICENSE.txt

-%description devel
-This is a C++ header-only library to provide an IEEE-754 conformant
-half-precision floating point type along with corresponding arithmetic
-operators, type conversions and common mathematical functions. It aims
-for both efficiency and ease of use, trying to accurately mimic the
-behaviour of the builtin floating point types at the best performance
-possible. It automatically uses and provides C++11 features when
-possible, but stays completely C++98-compatible when neccessary.
-
-%prep
-rm -rf %{name}-%{version}
-unzip -d %{name}-%{version} %{SOURCE0}
-cd %{name}-%{version}
-# change dos endings to unix
-sed -i "s|\r||g" include/half.hpp
-sed -i "s|\r||g" LICENSE.txt
-sed -i "s|\r||g" README.txt
-
-%install
-cd %{name}-%{version}
-mkdir -p %{buildroot}%{_includedir}
-install -m 644 include/half.hpp %{buildroot}%{_includedir}
-
-mkdir -p %{buildroot}%{_docdir}/%{name}/
-install -m 644 LICENSE.txt %{buildroot}%{_docdir}/%{name}/
-install -m 644 README.txt %{buildroot}%{_docdir}/%{name}/
-
-%files devel
-%dir %{_docdir}/%{name}
-%doc %{_docdir}/%{name}/README.txt
-%license %{_docdir}/%{name}/LICENSE.txt
-%{_includedir}/half.hpp
+%files
+%license LICENSE.txt
+%doc README.txt
+%{_includedir}/half/

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 2.2.0-1
- Import from upstream
+%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
@@ -1,204 +1,84 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
-%endif
+# 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 upstreamname hipcub
-%global rocm_release 7.1
-%global rocm_patch 1
-%global rocm_version %{rocm_release}.%{rocm_patch}
-
-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-
-# Compiler is hipcc, which is clang based:
-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//')
-# there is no debug package
-%global debug_package %{nil}
-
-# build test subpackage
 %bcond test 0
-
-# Option to test suite for testing on real HW:
-%bcond check 0
-
-%if %{with check} || %{with test}
+%if %{with test}
 %global build_test ON
 %else
 %global build_test OFF
 %endif

-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
+%global toolchain clang

-%global gpu_list %{rocm_gpu_list_default}
-%global _gpu_list gfx1100
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}

-Name:           hipcub%{pkg_suffix}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        2%{?dist}
-%else
+Name:           hipcub
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
-Summary:        ROCm port of CUDA CUB library
+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

-License:        MIT and BSD-3-Clause
-URL:            https://github.com/ROCm/rocm-libraries
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=%{build_test}

-%if %{with gitcommit}
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
-%else
-Source0:        %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
-%endif
-
-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocprim%{pkg_suffix}-static
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-
-%if %{with check} || %{with test}
-BuildRequires:  gtest-devel
-BuildRequires:  rocminfo%{pkg_suffix}
-%endif
-
-# Only headers, cmake infra but noarch confuses the libdir
-ExclusiveArch:  x86_64 riscv64
+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 wrapper library on top of rocPRIM or CUB. It enables developers
-to port a project using the CUB library to the HIP layer to run on AMD hardware.
-In the ROCm environment, hipCUB uses the rocPRIM library as the backend.
-
-%package devel
-Summary:        The %{upstreamname} development package
-Provides:       %{name}-static = %{version}-%{release}
-Requires:       rocprim%{pkg_suffix}-devel
-
-%description devel
-The %{upstreamname} development package.
+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:        Self-tests for %{name}
-Requires:       %{name}-devel%{?_isa} = %{version}-%{release}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
-Precompiled self-tests for %{name}
+%description    test
+%{summary}
 %endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/hipcub
-%else
-%autosetup -p1 -n %{upstreamname}
-%endif
+%prep -a
+# Fix cmake install lib directory
+sed -i -e 's/ROCM_INSTALL_LIBDIR lib/ROCM_INSTALL_LIBDIR %{_lib}/' \
+    cmake/ROCMExportTargetsHeaderOnly.cmake

-#
-# The ROCMExportTargetsHeaderOnly.cmake file
-# generates a files that reference the install location of other files
-# Make this change so they match
-sed -i -e 's/ROCM_INSTALL_LIBDIR lib/ROCM_INSTALL_LIBDIR %{pkg_libdir}/' cmake/ROCMExportTargetsHeaderOnly.cmake
+%install -a
+rm -f %{buildroot}/%{_datadir}/doc/hipcub/LICENSE.txt

-%build
-%if %{with gitcommit}
-cd projects/hipcub
-%endif
-
-
-%if %{with check}
-# Building all the gpu's does not make sense
-# Build only the first one, this only works well with rpmbuild.
-gpu=`rocm_agent_enumerator | head -n 1`
-%endif
-
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DBUILD_TEST=%{build_test} \
-%if %{with check}
-    -DAMDGPU_TARGETS=${gpu} \
-%else
-    -DAMDGPU_TARGETS=${gpu_list} \
-%endif
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DROCM_SYMLINK_LIBS=OFF
-
-%cmake_build
-
-%install
-%if %{with gitcommit}
-cd projects/hipcub
-%endif
-
-%cmake_install
-
-# Extra license
-rm -f %{buildroot}%{pkg_prefix}/share/doc/hipcub/LICENSE.txt
-
-%if %{with check}
-%check
-%ctest
-%endif
-
-%files devel
-%if %{with gitcommit}
-%doc projects/hipcub/README.md
-%license projects/hipcub/LICENSE.txt
-%else
+%files
 %doc README.md
 %license LICENSE.txt
-%endif
-%{pkg_prefix}/include/hipcub
-%{pkg_prefix}/%{pkg_libdir}/cmake/hipcub
+%{_includedir}/hipcub/
+%{_libdir}/cmake/hipcub/

 %if %{with test}
 %files test
-%{pkg_prefix}/bin/test_*
-%{pkg_prefix}/bin/hipcub/
+%{_bindir}/test_*
+%{_bindir}/hipcub/
 %endif

 %changelog
-* Mon Feb 9 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%autochangelog
@@ -1,224 +1,105 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
-%endif
+# 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 upstreamname hipfft
 %global rocm_release 7.1
 %global rocm_patch 1
 %global rocm_version %{rocm_release}.%{rocm_patch}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-%global hipfft_name hipfft%{pkg_suffix}
+# rocm stack builds with clang
+%global toolchain clang

-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//' -e 's/-flto=thin//' )
-
-%global _lto_cflags %{nil}
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%global __brp_check_rpaths %{nil}
-%else
-%global build_test OFF
-%endif
-
-# Option to test suite for testing on real HW:
-# May have to set gpu under test with
-# export HIP_VISIBLE_DEVICES=<num> - 0, 1 etc.
-%bcond check 0
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio" xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-Name:           %{hipfft_name}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        2%{?dist}
-%else
+Name:           hipfft
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
+Release:        %autorelease
 Summary:        ROCm FFT marshalling library
+Url:            https://github.com/ROCm/rocm-libraries
+VCS:            git:https://github.com/ROCm/hipFFT.git
 License:        MIT
-URL:            https://github.com/ROCm/rocm-libraries
-
-%if %{with gitcommit}
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
-%else
-Source0:        %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
-%endif
-
-# https://github.com/ROCm/rocm-libraries/issues/2400
+#!RemoteAsset:  sha256:f6f0352b5f9ffe53c88cea5fa40572eef0c0c1e2e50dce6f85d2c68e47afc63e
+Source:         %{url}/releases/download/rocm-%{version}/hipfft.tar.gz
 Patch1:         0001-hipfft-hipfftw-soversion.patch
+BuildSystem:    cmake

-ExclusiveArch:  x86_64 riscv64
+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:  llvm
-BuildRequires:  llvm-devel
-BuildRequires:  clang
-BuildRequires:  clang-devel
-BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
-BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocprim%{pkg_suffix}-static
-BuildRequires:  rocfft%{pkg_suffix}-devel
-
-%if %{with test}
 BuildRequires:  boost-devel
-BuildRequires:  fftw-devel
-BuildRequires:  hiprand%{pkg_suffix}-devel
-BuildRequires:  rocrand%{pkg_suffix}-devel
-
-BuildRequires:  gtest-devel
-%endif
-
-Provides:       hipfft%{pkg_suffix} = %{version}-%{release}
+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 an FFT marshalling library. Currently, hipFFT supports
-the rocFFT backends
+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.

-hipFFT exports an interface that does not require the client to
-change, regardless of the chosen backend. It sits between the
-application and the backend FFT library, marshalling inputs into
-the backend and results back to the application.
-
-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
-Summary:        Libraries and headers for %{name}
+%package        devel
+Summary:        The hipFFT development package
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Provides:       hipfft%{pkg_suffix}-devel = %{version}-%{release}
+Requires:       cmake(rocfft)

-%description devel
-%{summary}
+%description    devel
+The hipFFT development package.

-%if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
-%endif
-
-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/hipfft
-%else
-%autosetup -p1 -n %{upstreamname}
-%endif

+%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
+sed -i -e 's@find_package( HIP MODULE REQUIRED )@find_package( HIP REQUIRED )@' \
+    clients/tests/CMakeLists.txt

-%build
-%if %{with gitcommit}
-cd projects/hipfft
-%endif
+%install -a
+rm -f %{buildroot}/%{_datadir}/doc/hipfft/LICENSE.md

-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DCMAKE_BUILD_TYPE=%{build_type} \
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DCMAKE_SKIP_RPATH=ON \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DAMDGPU_TARGETS=%{rocm_gpu_list_default} \
-    -DBUILD_CLIENTS_TESTS=%{build_test} \
-    -DBUILD_CLIENTS_TESTS_OPENMP=OFF \
-    -DROCM_SYMLINK_LIBS=OFF \
-    -DHIP_PLATFORM=amd
-
-%cmake_build
-
-
-%install
-%if %{with gitcommit}
-cd projects/hipfft
-%endif
-%cmake_install
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/hipfft/LICENSE.md
-
-%check
 %if %{with test}
-%if %{with check}
-export LD_LIBRARY_PATH=%{_vpath_builddir}/library:$LD_LIBRARY_PATH
-%{_vpath_builddir}/clients/staging/hipfft-test
-%endif
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH
 %endif

 %files
-%if %{with gitcommit}
-%license projects/hipfft/LICENSE.md
-%doc projects/hipfft/README.md
-%else
-%license LICENSE.md
 %doc README.md
-%endif
-
-%{pkg_prefix}/%{pkg_libdir}/libhipfft.so.0{,.*}
-%{pkg_prefix}/%{pkg_libdir}/libhipfftw.so.0{,.*}
+%license LICENSE.md
+%{_libdir}/libhipfft.so.0{,.*}
+%{_libdir}/libhipfftw.so.0{,.*}

 %files devel
-%{pkg_prefix}/include/hipfft/
-%{pkg_prefix}/%{pkg_libdir}/libhipfft.so
-%{pkg_prefix}/%{pkg_libdir}/libhipfftw.so
-%{pkg_prefix}/%{pkg_libdir}/cmake/hipfft/
+%{_includedir}/hipfft/
+%{_libdir}/cmake/hipfft/
+%{_libdir}/libhipfft.so
+%{_libdir}/libhipfftw.so

-%if %{with test}
 %files test
-%{pkg_prefix}/bin/hipfft-test
-%endif
+%{_bindir}/hipfft-test

 %changelog
-* Mon Feb 9 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%autochangelog
@@ -1,107 +1,52 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
-%endif
+# 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 upstreamname hiprand
 %global rocm_release 7.1
 %global rocm_patch 1
 %global rocm_version %{rocm_release}.%{rocm_patch}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-%global hiprand_name hiprand%{pkg_suffix}
+# rocm stack builds with clang
+%global toolchain clang

-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//')
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%else
-%global build_test OFF
-%endif
-
-# Option to test suite for testing on real HW:
-%bcond check 0
-# For docs
-%bcond doc 0
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-Name:           %{hiprand_name}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        3%{?dist}
-%else
+Name:           hiprand
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
+Release:        %autorelease
 Summary:        HIP random number generator
 License:        MIT AND BSD-3-Clause
-URL:            https://github.com/ROCm/rocm-libraries
-%if %{with gitcommit}
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
-%else
-Source0:        %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
-%endif
+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:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-BuildRequires:  rocrand%{pkg_suffix}-devel
-
-%if %{with test}
-BuildRequires:  gtest-devel
-%endif
-
-%if %{with doc}
-BuildRequires:  doxygen
-%endif
-
-Provides:       hiprand%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64
+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
@@ -110,113 +55,48 @@ 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.

-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
+%package        devel
 Summary:        The hipRAND development package
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Requires:       rocrand%{pkg_suffix}-devel
-Provides:       hiprand%{pkg_suffix}-devel = %{version}-%{release}
+Requires:       cmake(rocrand)

-%description devel
+%description    devel
 The hipRAND development package.

-%if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
-%endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/hiprand
-%else
-%autosetup -p1 -n %{upstreamname}
-%endif
-
-#Remove RPATH:
+%prep -a
+# Remove RPATH
 sed -i '/INSTALL_RPATH/d' CMakeLists.txt

-# On Tumbleweed Q2,2025
-# /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.
-# https://github.com/ROCm/hipRAND/issues/222
-# Convert the c++11's to c++14
-sed -i -e 's@set(CMAKE_CXX_STANDARD 11)@set(CMAKE_CXX_STANDARD 14)@' {,test/package/}CMakeLists.txt
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hiprand/LICENSE.md
+rm -f %{buildroot}%{_bindir}/hipRAND/CTestTestfile.cmake

-%build
-%if %{with gitcommit}
-cd projects/hiprand
-%endif
-
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DCMAKE_BUILD_TYPE=%{build_type} \
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DCMAKE_SKIP_RPATH=ON \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DAMDGPU_TARGETS=%{rocm_gpu_list_default} \
-    -DBUILD_TEST=%{build_test} \
-    -DROCM_SYMLINK_LIBS=OFF
-
-%cmake_build
-
-%install
-%if %{with gitcommit}
-cd projects/hiprand
-%endif
-%cmake_install
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/hiprand/LICENSE.md
-rm -f %{buildroot}%{pkg_prefix}/bin/hipRAND/CTestTestfile.cmake
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH

+%if %{without run_test}
 %check
-%if %{with test}
-%if %{with check}
-
-%ctest
-%endif
 %endif

 %files
-%if %{with gitcommit}
-%doc projects/hiprand/README.md
-%license projects/hiprand/LICENSE.md
-%else
 %doc README.md
 %license LICENSE.md
-%endif
-%if %{with debug}
-%{pkg_prefix}/%{pkg_libdir}/libhiprand-d.so.1{,.*}
-%else
-%{pkg_prefix}/%{pkg_libdir}/libhiprand.so.1{,.*}
-%endif
+%{_libdir}/libhiprand.so.1{,.*}

 %files devel
-%{pkg_prefix}/include/hiprand/
-%if %{with debug}
-%{pkg_prefix}/%{pkg_libdir}/libhiprand-d.so
-%else
-%{pkg_prefix}/%{pkg_libdir}/libhiprand.so
-%endif
-%{pkg_prefix}/%{pkg_libdir}/cmake/hiprand/
+%{_includedir}/hiprand/
+%{_libdir}/cmake/hiprand/
+%{_libdir}/libhiprand.so

-%if %{with test}
 %files test
-%{pkg_prefix}/bin/test*
-%endif
+%{_bindir}/test*

 %changelog
-* Mon Feb 9 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%autochangelog
@@ -1,206 +1,124 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
+# 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

-%global upstreamname hipSOLVER
+# 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}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-%global hipsolver_name hipsolver%{pkg_suffix}
+# rocm stack builds with clang
+%global toolchain clang

-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//')
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%else
-%global build_test OFF
-%endif
-
-# gfortran and clang rpm macros do not mix
+# Fortran is only used in testing
+# clang and gfortran fedora toolchain args do not mix
 %global build_fflags %{nil}

-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-Name:           %{hipsolver_name}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        3%{?dist}
-%else
+Name:           hipsolver
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
-Summary:        ROCm SOLVER marshalling library
+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

-%if %{with gitcommit}
-Url:            https://github.com/ROCm/rocm-libraries
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
-%else
-Url:            https://github.com/ROCm/%{upstreamname}
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz#/%{upstreamname}-%{rocm_version}.tar.gz
-%endif
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DBUILD_CLIENTS_TESTS=%{cmake_test}
+BuildOption(conf):  -DBUILD_CLIENTS_BENCHMARKS=%{cmake_test}

-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
+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:  rocblas%{pkg_suffix}-devel
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocsolver%{pkg_suffix}-devel
-BuildRequires:  rocsparse%{pkg_suffix}-devel
-
-%if %{with test}
-BuildRequires:  gtest-devel
-BuildRequires:  hipsparse%{pkg_suffix}-devel
-BuildRequires:  blas-static
-BuildRequires:  lapack-static
+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

-Provides:       hipsolver%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64
-
 %description
-hipSOLVER is a LAPACK marshalling library, with multiple supported
-backends. It sits between the application and a 'worker'
-LAPACK library, marshalling inputs into the backend library and
-marshalling results back to the application. hipSOLVER exports an
-interface that does not require the client to change, regardless
-of the chosen backend. Currently, hipSOLVER supports rocSOLVER
-and cuSOLVER as backends.
+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.

-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
-Summary:        Libraries and headers for %{name}
+%package        devel
+Summary:        The hipSOLVER development package
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Provides:       hipsolver%{pkg_suffix}-devel = %{version}-%{release}
+Requires:       cmake(rocblas)
+Requires:       cmake(rocsolver)
+Requires:       cmake(rocsparse)

-%description devel
-%{summary}
+%description    devel
+The hipSOLVER development package.

-%if %{with test}
-%package test
+%if %{with build_test}
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
 %endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/hipsolver
-%patch -P1 -p1
-%else
-%autosetup -p1 -n %{upstreamname}-rocm-%{version}
+%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

-%build
-%if %{with gitcommit}
-cd projects/hipsolver
-%endif
-
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DCMAKE_BUILD_TYPE=%{build_type} \
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DCMAKE_SKIP_RPATH=ON \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DROCM_SYMLINK_LIBS=OFF \
-    -DHIP_PLATFORM=amd \
-    -DAMDGPU_TARGETS=%{rocm_gpu_list_default} \
-    -DBUILD_CLIENTS_TESTS=%{build_test}
-
-%cmake_build
-
-%install
-%if %{with gitcommit}
-cd projects/hipsolver
-%endif
-
-%cmake_install
-
-# Extra license
-rm -f %{buildroot}%{pkg_prefix}/share/doc/hipsolver/LICENSE.md
-
 %files
-%if %{with gitcommit}
-%doc projects/hipsolver/README.md
-%license projects/hipsolver/LICENSE.md
-%else
 %doc README.md
 %license LICENSE.md
-%endif
-
-%{pkg_prefix}/%{pkg_libdir}/libhipsolver.so.1{,.*}
-%{pkg_prefix}/%{pkg_libdir}/libhipsolver_fortran.so.1{,.*}
+%{_libdir}/libhipsolver.so.1{,.*}
+%{_libdir}/libhipsolver_fortran.so.1{,.*}

 %files devel
-%{pkg_prefix}/include/hipsolver/
-%{pkg_prefix}/%{pkg_libdir}/libhipsolver.so
-%{pkg_prefix}/%{pkg_libdir}/libhipsolver_fortran.so
-%{pkg_prefix}/%{pkg_libdir}/cmake/hipsolver/
+%{_includedir}/hipsolver/
+%{_libdir}/libhipsolver.so
+%{_libdir}/libhipsolver_fortran.so
+%{_libdir}/cmake/hipsolver/

-%if %{with test}
+%if %{with build_test}
 %files test
-%{pkg_prefix}/share/hipsolver/
-%{pkg_prefix}/bin/hipsolver*
+%{_datadir}/hipsolver/
+%{_bindir}/hipsolver*
 %endif

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%autochangelog
@@ -1,56 +0,0 @@
-From 4105bd51e654df8b40c53eb4b4abe7a3415783eb Mon Sep 17 00:00:00 2001
-From: Tom Rix <Tom.Rix@amd.com>
-Date: Fri, 31 Oct 2025 07:28:05 -0700
-Subject: [PATCH] hipsparse change test download dir
-
---
- clients/tests/CMakeLists.txt | 22 ++++++++++++----------
- 1 file changed, 12 insertions(+), 10 deletions(-)
-
-diff --git a/clients/tests/CMakeLists.txt b/clients/tests/CMakeLists.txt
-index 15a7c21f4c30..69beb9da6208 100644
--- a/clients/tests/CMakeLists.txt
-+++ b/clients/tests/CMakeLists.txt
-@@ -27,11 +27,13 @@ find_package(GTest REQUIRED)
- #
- # Client matrices.
- #
-+option(CMAKE_MATRICES_DIR "Directory containing test matrices" ${PROJECT_BINARY_DIR}/matrices)
-+
- if(NOT EXISTS "${CMAKE_MATRICES_DIR}")
-   #
-   # Download.
-   #
-  set(CMAKE_MATRICES_DIR ${PROJECT_BINARY_DIR}/matrices CACHE STRING "Matrices directory.")
-+
-   file(MAKE_DIRECTORY ${CMAKE_MATRICES_DIR})
- 
-   if(NOT TARGET hipsparse)
-@@ -46,15 +48,15 @@ else()
-   #
-   # Copy.
-   #
-  if(NOT CMAKE_MATRICES_DIR STREQUAL "${PROJECT_BINARY_DIR}/matrices")
-    message("Copy matrix files from ${CMAKE_MATRICES_DIR} to ${PROJECT_BINARY_DIR}/matrices")
-
-    execute_process(COMMAND cp -r ${CMAKE_MATRICES_DIR} ${PROJECT_BINARY_DIR}/matrices RESULT_VARIABLE STATUS WORKING_DIRECTORY ${CMAKE_MATRICES_DIR})
-
-    if(STATUS AND NOT STATUS EQUAL 0)
-      message(FATAL_ERROR "Failed to copy matrix .bin files, aborting.")
-    endif()
-  endif()
-+  #  if(NOT CMAKE_MATRICES_DIR STREQUAL "${PROJECT_BINARY_DIR}/matrices")
-+  #    message("Copy matrix files from ${CMAKE_MATRICES_DIR} to ${PROJECT_BINARY_DIR}/matrices")
-+  #
-+  #    execute_process(COMMAND cp -r ${CMAKE_MATRICES_DIR} ${PROJECT_BINARY_DIR}/matrices RESULT_VARIABLE STATUS WORKING_DIRECTORY ${CMAKE_MATRICES_DIR})
-+  #
-+  #    if(STATUS AND NOT STATUS EQUAL 0)
-+  #      message(FATAL_ERROR "Failed to copy matrix .bin files, aborting.")
-+  #    endif()
-+  #  endif()
- 
- endif()
- 
-- 
-2.51.0
-
@@ -1,129 +1,66 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
+# 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

-%global upstreamname hipsparse
+# 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}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-%global hipsparse_name hipsparse%{pkg_suffix}
-
-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//')
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# export an llvm compilation database
-# Useful for input for other llvm tools
-%bcond export 0
-%if %{with export}
-%global build_compile_db ON
-%else
-%global build_compile_db OFF
-%endif
-
-# downloads tests, use mock --enable-network
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%global __brp_check_rpaths %{nil}
-%else
-%global build_test OFF
-%endif
-
-%bcond check 0
-
-# gfortran and clang rpm macros do not mix
-%global build_fflags %{nil}
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-Name:           %{hipsparse_name}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        1%{?dist}
-%else
+Name:           hipsparse
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
+Release:        %autorelease
 Summary:        ROCm SPARSE marshalling library
 License:        MIT
-URL:            https://github.com/ROCm/rocm-libraries
+Url:            https://github.com/ROCm/hipSPARSE
+#!RemoteAsset:  sha256:b001834d8e65c3878d1a69d08803d5b6ce4fe623e78099fe51cb146d0ffa10e7
+Source:         %{url}/archive/rocm-%{version}.tar.gz
+BuildSystem:    cmake

-%if %{with gitcommit}
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
-%else
-Source0:        %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
-%endif
+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}

-Patch1:         0001-hipsparse-change-test-download-dir.patch
-
-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
+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:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocprim%{pkg_suffix}-static
-BuildRequires:  rocsparse%{pkg_suffix}-devel
-
-%if %{with test}
-BuildRequires:  gtest-devel
-BuildRequires:  rocblas%{pkg_suffix}-devel
-%endif
-
-%if %{with check}
-%if %{with export}
-BuildRequires:  cppcheck
-BuildRequires:  cppcheck-htmlreport
-BuildRequires:  rocm-clang-analyzer%{pkg_suffix}
-BuildRequires:  rocm-clang-tools-extra%{pkg_suffix}
-%endif
-%endif
-
-Provides:       hipsparse%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64
+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
@@ -135,126 +72,57 @@ require the client to change, regardless of the chosen
 backend. Currently, hipSPARSE supports rocSPARSE and
 cuSPARSE backends.

-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
-Summary:        Libraries and headers for %{name}
+%package        benchmark
+Summary:        Benchmark for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Provides:       hipsparse%{pkg_suffix}-devel = %{version}-%{release}

-%description devel
+%description    benchmark
 %{summary}

-%if %{with test}
-%package test
+%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
+%description    test
 %{summary}
 %endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/hiprand
-%patch -P1 -p1
-%else
-%autosetup -p1 -n %{upstreamname}
-%endif
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/hipsparse/LICENSE.md

-# A better default for the matrices dir
-sed -i -e 's@hipsparse_exepath() + "../matrices/"@"%{pkg_prefix}/share/hipsparse/matrices/"@' clients/include/utility.hpp
+%check -p
+export LD_LIBRARY_PATH=$PWD/%{__cmake_builddir}/library:$LD_LIBRARY_PATH

-%build
-%if %{with gitcommit}
-cd projects/hipsparse
-%endif
-
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_EXPORT_COMPILE_COMMANDS=%{build_compile_db} \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DCMAKE_BUILD_TYPE=%build_type \
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DCMAKE_SKIP_RPATH=ON \
-    -DROCM_SYMLINK_LIBS=OFF \
-    -DHIP_PLATFORM=amd \
-    -DGPU_TARGETS=%{rocm_gpu_list_default} \
-    -DBUILD_CLIENTS_BENCHMARKS=%{build_test} \
-    -DBUILD_CLIENTS_SAMPLES=OFF \
-    -DBUILD_CLIENTS_TESTS=%{build_test} \
-    -DBUILD_CLIENTS_TESTS_OPENMP=OFF \
-    -DCMAKE_MATRICES_DIR=%{_builddir}/hipsparse-test-matrices/ \
-    -DBUILD_FORTRAN_CLIENTS=OFF
-
-%cmake_build
-
-%if %{with check}
+%if %{without run_test}
 %check
-%if %{with export}
-json=`find . -name 'compile_commands.json'`
-json=`realpath $json`
-json_dir=`dirname $json`
-if [ -f ${json} ]; then
-    jobs=`nproc`
-    export PATH=%{rocmllvm_bindir}:$PATH
-    output=/tmp/%{name}-tidy/
-    mkdir -p ${output}
-    # Use echo to consume tidy's error code
-    %{rocmllvm_bindir}/run-clang-tidy -p ${json_dir} &> ${output}/tidy.log || echo "ran clang-tidy"
-    
-    output=/tmp/%{name}-cppcheck/
-    mkdir -p ${output}
-    cppcheck --project=${json} -j ${jobs} --std=c++17 --safety --output-file=${output}/cppcheck.txt
-    cppcheck --project=${json} -j ${jobs} --std=c++17 --safety --xml --output-file=${output}/cppcheck.xml
-    cppcheck-htmlreport --file=${output}/cppcheck.xml --report-dir=${output}
-fi
-%endif
-%endif
-
-%install
-%if %{with gitcommit}
-cd projects/hipsparse
-%endif
-
-%cmake_install
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/hipsparse/LICENSE.md
-
-%if %{with test}
-mkdir -p %{buildroot}/%{pkg_prefix}/share/hipsparse/matrices
-install -pm 644 %{_builddir}/%{name}-test-matrices/* %{buildroot}/%{pkg_prefix}/share/hipsparse/matrices
 %endif

 %files
-%if %{with gitcommit}
-%doc projects/hipsparse/README.md
-%license projects/hipsparse/LICENSE.md
-%else
 %doc README.md
 %license LICENSE.md
-%endif
+%{_libdir}/libhipsparse.so.4{,.*}

-%{pkg_prefix}/%{pkg_libdir}/libhipsparse.so.4{,.*}
+%files benchmark
+%{_bindir}/hipsparse-bench

 %files devel
-%{pkg_prefix}/include/hipsparse/
-%{pkg_prefix}/%{pkg_libdir}/libhipsparse.so
-%{pkg_prefix}/%{pkg_libdir}/cmake/hipsparse/
+%{_includedir}/hipsparse/
+%{_libdir}/cmake/hipsparse/
+%{_libdir}/libhipsparse.so

-%if %{with test}
+%if %{with build_test}
 %files test
-%{pkg_prefix}/bin/hipsparse*
-%{pkg_prefix}/share/hipsparse/
+%{_bindir}/hipsparse*
+%{_datadir}/hipsparse/
 %endif

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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
@@ -1,28 +0,0 @@
-From 19a83736d99acc42308f8003094c27b920d1c285 Mon Sep 17 00:00:00 2001
-From: Tom Rix <trix@redhat.com>
-Date: Sun, 7 Jul 2024 05:07:39 -0600
-Subject: [PATCH] Prepare magma cmake for fedora
-
-Need to version *.so'
-
-Signed-off-by: Tom Rix <trix@redhat.com>
---
- CMakeLists.txt | 2 ++
- 1 file changed, 2 insertions(+)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 0c9b41c7d7fb..06e401e1199f 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -760,6 +760,8 @@ foreach( TEST ${sparse_testing_all} )
- endforeach()
- add_custom_target( sparse-testing DEPENDS ${sparse-testing} )
- 
-+SET_TARGET_PROPERTIES(magma PROPERTIES VERSION magma_VERSION SOVERSION magma_VERSION)
-+SET_TARGET_PROPERTIES(magma_sparse PROPERTIES VERSION magma_VERSION SOVERSION magma_VERSION)
- 
- # ----------------------------------------
- # what to install
-- 
-2.45.1
-
@@ -1,754 +0,0 @@
-From 4afc1c3847ef926d7a29e18c6bf79cc4623805a1 Mon Sep 17 00:00:00 2001
-From: Tom Rix <Tom.Rix@amd.com>
-Date: Fri, 3 Oct 2025 14:47:42 -0700
-Subject: [PATCH] magma ROCm 7 changes
-
-From Jeff Daily's PR
-https://github.com/jeffdaily/magma/commit/1b966b72402e3f37ebd462f3d7e019e669e510ff
-
-Signed-off-by: Tom Rix <Tom.Rix@amd.com>
---
- include/magma_types.h                       |  5 ----
- interface_cuda/blas_h_v2.cpp                | 16 +++++------
- magmablas/zgemm_batched.cpp                 |  4 +--
- magmablas/zgetf2_kernels.cu                 |  2 +-
- make.inc-examples/make.inc.hip-gcc-mkl      |  6 ----
- make.inc-examples/make.inc.hip-gcc-openblas |  6 ----
- sparse/blas/magma_z_blaswrapper.cpp         |  9 ------
- sparse/blas/magma_zmatrixtools_gpu.cu       |  5 ----
- sparse/blas/magma_ztrisolve.cpp             |  9 ------
- sparse/blas/zilu.cpp                        |  9 ------
- sparse/blas/zmergecg.cu                     | 10 -------
- sparse/control/magma_zmconvert.cpp          |  5 ----
- sparse/control/magma_zmtranspose.cpp        |  5 ----
- sparse/src/zcustomic.cpp                    |  5 ----
- sparse/src/zcustomilu.cpp                   |  5 ----
- sparse/src/zparict.cpp                      |  6 ----
- sparse/src/zparilut.cpp                     |  5 ----
- sparse/testing/testing_zspmm.cpp            |  9 ------
- sparse/testing/testing_zspmv.cpp            |  9 ------
- src/shpotrf_gpu.cpp                         |  8 +++---
- src/xhsgetrf_gpu.cpp                        | 32 ++++++++++-----------
- src/xshgetrf_gpu.cpp                        | 32 ++++++++++-----------
- testing/testing_sgemm_fp16.cpp              | 16 +++++------
- testing/testing_zgemm_batched.cpp           | 20 ++++++-------
- testing/testing_zgemv_batched.cpp           | 20 ++++++-------
- testing/testing_zgeqrf_batched.cpp          |  4 +--
- testing/testing_zgetrf_batched.cpp          |  2 +-
- testing/testing_ztrsm_batched.cpp           |  6 ++--
- tools/hipify-perl                           | 17 ++++-------
- tools/magmasubs.py                          |  3 +-
- 30 files changed, 87 insertions(+), 203 deletions(-)
-
-diff --git a/include/magma_types.h b/include/magma_types.h
-index 7f9c333c2643..2c03f5f8cc34 100644
--- a/include/magma_types.h
-+++ b/include/magma_types.h
-@@ -227,8 +227,6 @@ typedef double real_Double_t;
- 
-     /* double complex */
- 
-    //typedef hipblasDoubleComplex magmaDoubleComplex;
-
-     /* simple double complex definition that should be binary compatible with hipBLAS */
-     typedef struct {
- 
-@@ -275,9 +273,6 @@ typedef double real_Double_t;
- 
-     /* float complex */
- 
-    //typedef hipComplex magmaFloatComplex;
-    //typedef hipblasComplex magmaFloatComplex;
-
-     /* basic definition of float complex that should be binary compatible with hipBLAS */
-     typedef struct {
- 
-diff --git a/interface_cuda/blas_h_v2.cpp b/interface_cuda/blas_h_v2.cpp
-index 2f61a1d6f5e2..56683fe294b0 100644
--- a/interface_cuda/blas_h_v2.cpp
-+++ b/interface_cuda/blas_h_v2.cpp
-@@ -118,10 +118,10 @@ magma_hgemm(
- 		      hipblas_trans_const( transA ),
- 		      hipblas_trans_const( transB ),
- 		      int(m), int(n), int(k),
-		      (void*)&alpha, (void*)dA, HIPBLAS_R_16F, int(ldda),
-		      (void*)dB, HIPBLAS_R_16F, int(lddb),
-		      (void *)&beta,  (void*)dC, HIPBLAS_R_16F, int(lddc),
-		      HIPBLAS_R_16F,
-+		      (void*)&alpha, (void*)dA, HIP_R_16F, int(ldda),
-+		      (void*)dB, HIP_R_16F, int(lddb),
-+		      (void *)&beta,  (void*)dC, HIP_R_16F, int(lddc),
-+		      HIPBLAS_COMPUTE_16F,
- 		      HIPBLAS_GEMM_DEFAULT);
-     }
-     else {
-@@ -151,10 +151,10 @@ magma_hgemmx(
- 		      hipblas_trans_const( transA ),
- 		      hipblas_trans_const( transB ),
- 		      int(m), int(n), int(k),
-		      (void*)&alpha, (void*)dA, HIPBLAS_R_16F, int(ldda),
-                                     (void*)dB, HIPBLAS_R_16F, int(lddb),
-		      (void*)&beta,  (void*)dC, HIPBLAS_R_32F, int(lddc),
-		      HIPBLAS_R_32F,
-+		      (void*)&alpha, (void*)dA, HIP_R_16F, int(ldda),
-+                                     (void*)dB, HIP_R_16F, int(lddb),
-+		      (void*)&beta,  (void*)dC, HIP_R_32F, int(lddc),
-+		      HIPBLAS_COMPUTE_32F,
- 		      HIPBLAS_GEMM_DEFAULT);
-     }
-     else {
-diff --git a/magmablas/zgemm_batched.cpp b/magmablas/zgemm_batched.cpp
-index 2b4ea3ba1750..86a82971da0d 100644
--- a/magmablas/zgemm_batched.cpp
-+++ b/magmablas/zgemm_batched.cpp
-@@ -28,13 +28,13 @@
-  *       */
- #ifdef PRECISION_z
-   #ifdef MAGMA_HAVE_HIP
-    typedef hipblasDoubleComplex BackendFloat_t;
-+    typedef hipDoubleComplex BackendFloat_t;
-   #else
-     typedef cuDoubleComplex BackendFloat_t;
-   #endif
- #elif defined(PRECISION_c)
-   #ifdef MAGMA_HAVE_HIP
-    typedef hipblasComplex BackendFloat_t;
-+    typedef hipComplex BackendFloat_t;
-   #else
-     typedef cuFloatComplex BackendFloat_t;
-   #endif
-diff --git a/magmablas/zgetf2_kernels.cu b/magmablas/zgetf2_kernels.cu
-index db8e7058fd11..3ff73755dda4 100644
--- a/magmablas/zgetf2_kernels.cu
-+++ b/magmablas/zgetf2_kernels.cu
-@@ -211,7 +211,7 @@ magma_izamax_native(
-         hipblasGetPointerMode(queue->hipblas_handle(), &ptr_mode);
-         hipblasSetPointerMode(queue->hipblas_handle(), CUBLAS_POINTER_MODE_DEVICE);
- 
-        hipblasIzamax(queue->hipblas_handle(), length, (const hipblasDoubleComplex*)x, 1, (int*)(ipiv));
-+        hipblasIzamax(queue->hipblas_handle(), length, (const hipDoubleComplex*)x, 1, (int*)(ipiv));
-         magma_zpivcast<<< 1, 1, 0, queue->cuda_stream() >>>( ipiv );
- 
-         hipblasSetPointerMode(queue->hipblas_handle(), ptr_mode);
-diff --git a/make.inc-examples/make.inc.hip-gcc-mkl b/make.inc-examples/make.inc.hip-gcc-mkl
-index 4a7809e208f5..349835bbe2bd 100644
--- a/make.inc-examples/make.inc.hip-gcc-mkl
-+++ b/make.inc-examples/make.inc.hip-gcc-mkl
-@@ -112,12 +112,6 @@ ifeq ($(BACKEND),cuda)
-     DEVCCFLAGS += -Xcompiler "$(FPIC)" -Xcompiler "$(FOPENMP)" -std=c++11
- else ifeq ($(BACKEND),hip)
-     DEVCCFLAGS += $(FPIC) $(FOPENMP) -std=c++11
-    # check for older versions of ROCM
-    ifeq ($(shell hipconfig --version | cut -b -3),3.0)
-        $(info Building with HIP 3.0)
-        # they don't have hipblasComplex yet, so replace it manually
-        DEVCCFLAGS += -DhipblasComplex=hipComplex -DhipblasDoubleComplex=hipDoubleComplex
-    endif
- endif
- 
- 
-diff --git a/make.inc-examples/make.inc.hip-gcc-openblas b/make.inc-examples/make.inc.hip-gcc-openblas
-index 2c6be03ea343..21e6d83396ce 100644
--- a/make.inc-examples/make.inc.hip-gcc-openblas
-+++ b/make.inc-examples/make.inc.hip-gcc-openblas
-@@ -116,12 +116,6 @@ ifeq ($(BACKEND),cuda)
-     DEVCCFLAGS += -Xcompiler "$(FPIC)" -Xcompiler "$(FOPENMP)" -std=c++11
- else ifeq ($(BACKEND),hip)
-     DEVCCFLAGS += $(FPIC) $(FOPENMP) -std=c++11
-    # check for older versions of ROCM
-    ifeq ($(shell hipconfig --version | cut -b -3),3.0)
-        $(info Building with HIP 3.0)
-        # they don't have hipblasComplex yet, so replace it manually
-        DEVCCFLAGS += -DhipblasComplex=hipComplex -DhipblasDoubleComplex=hipDoubleComplex
-    endif
- endif
- 
- 
-diff --git a/sparse/blas/magma_z_blaswrapper.cpp b/sparse/blas/magma_z_blaswrapper.cpp
-index 1e3d3d887df7..e70618e3c812 100644
--- a/sparse/blas/magma_z_blaswrapper.cpp
-+++ b/sparse/blas/magma_z_blaswrapper.cpp
-@@ -13,15 +13,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-  #elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
- #if CUDA_VERSION >= 12000
-   #define CUSPARSE_CSRMV_ALG2 CUSPARSE_SPMV_CSR_ALG2
-   #define CUSPARSE_CSRMV_ALG1 CUSPARSE_SPMV_CSR_ALG1
-diff --git a/sparse/blas/magma_zmatrixtools_gpu.cu b/sparse/blas/magma_zmatrixtools_gpu.cu
-index dcb8952d0e04..ba801300d50e 100644
--- a/sparse/blas/magma_zmatrixtools_gpu.cu
-+++ b/sparse/blas/magma_zmatrixtools_gpu.cu
-@@ -14,11 +14,6 @@
- 
- #define SWAP(a, b)  { tmp = a; a = b; b = tmp; }
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- 
- 
- __global__ void 
-diff --git a/sparse/blas/magma_ztrisolve.cpp b/sparse/blas/magma_ztrisolve.cpp
-index f51ee65b3462..29aa54b894f3 100644
--- a/sparse/blas/magma_ztrisolve.cpp
-+++ b/sparse/blas/magma_ztrisolve.cpp
-@@ -13,15 +13,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-  #elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
- magma_int_t magma_ztrisolve_analysis(magma_z_matrix M, magma_solve_info_t *solve_info, bool upper_triangular, bool unit_diagonal, bool transpose, magma_queue_t queue)
- {
-     magma_int_t info = 0;
-diff --git a/sparse/blas/zilu.cpp b/sparse/blas/zilu.cpp
-index c3f5709bb678..4150de259830 100644
--- a/sparse/blas/zilu.cpp
-+++ b/sparse/blas/zilu.cpp
-@@ -16,15 +16,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-  #elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
- #if CUDA_VERSION >= 12000
-    #define cusparseCreateCsrsm2Info(info)
-    #define cusparseDestroyCsrsm2Info(info)
-diff --git a/sparse/blas/zmergecg.cu b/sparse/blas/zmergecg.cu
-index 7b472fa8591e..44d4b8720cd3 100644
--- a/sparse/blas/zmergecg.cu
-+++ b/sparse/blas/zmergecg.cu
-@@ -13,16 +13,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-  #elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
-
- #define BLOCK_SIZE 512
- 
- #if CUDA_VERSION >= 12000
-diff --git a/sparse/control/magma_zmconvert.cpp b/sparse/control/magma_zmconvert.cpp
-index 15e9a3d9a52c..83948bb3a198 100644
--- a/sparse/control/magma_zmconvert.cpp
-+++ b/sparse/control/magma_zmconvert.cpp
-@@ -13,11 +13,6 @@
- #include <cuda.h>  // for CUDA_VERSION
- 
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- 
- // todo: check if we need buf later
- #if CUDA_VERSION >= 11000
-diff --git a/sparse/control/magma_zmtranspose.cpp b/sparse/control/magma_zmtranspose.cpp
-index 60b91f1d159b..9de0193495f5 100644
--- a/sparse/control/magma_zmtranspose.cpp
-+++ b/sparse/control/magma_zmtranspose.cpp
-@@ -14,11 +14,6 @@
- 
- #include <cuda.h>  // for CUDA_VERSION
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- // todo: check if we need buf later
- #if CUDA_VERSION >= 11000
- #define cusparseZcsr2csc(handle, m, n, nnz, valA, rowA, colA, valB, colB, rowB,                \
-diff --git a/sparse/src/zcustomic.cpp b/sparse/src/zcustomic.cpp
-index 1859cd29fe7c..5214769be4f4 100644
--- a/sparse/src/zcustomic.cpp
-+++ b/sparse/src/zcustomic.cpp
-@@ -16,11 +16,6 @@
- 
- #define COMPLEX
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- /**
-     Purpose
-     -------
-diff --git a/sparse/src/zcustomilu.cpp b/sparse/src/zcustomilu.cpp
-index c118125f3e37..2ceb8844a537 100644
--- a/sparse/src/zcustomilu.cpp
-+++ b/sparse/src/zcustomilu.cpp
-@@ -16,11 +16,6 @@
- 
- #define COMPLEX
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- /**
-     Purpose
-     -------
-diff --git a/sparse/src/zparict.cpp b/sparse/src/zparict.cpp
-index d1e72ea7683c..8337fd69c8b3 100644
--- a/sparse/src/zparict.cpp
-+++ b/sparse/src/zparict.cpp
-@@ -21,12 +21,6 @@
- #define PRECISION_z
- 
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
-
- /***************************************************************************//**
-     Purpose
-     -------
-diff --git a/sparse/src/zparilut.cpp b/sparse/src/zparilut.cpp
-index 55caecd13e3b..747e1620df63 100644
--- a/sparse/src/zparilut.cpp
-+++ b/sparse/src/zparilut.cpp
-@@ -20,11 +20,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#ifdef MAGMA_HAVE_HIP
-  #define hipblasDoubleComplex hipDoubleComplex
-#endif
-
- /***************************************************************************//**
-     Purpose
-     -------
-diff --git a/sparse/testing/testing_zspmm.cpp b/sparse/testing/testing_zspmm.cpp
-index de21c66c8c57..2682af900f53 100644
--- a/sparse/testing/testing_zspmm.cpp
-+++ b/sparse/testing/testing_zspmm.cpp
-@@ -70,15 +70,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-#elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
- /* ////////////////////////////////////////////////////////////////////////////
-    -- testing sparse matrix vector product
- */
-diff --git a/sparse/testing/testing_zspmv.cpp b/sparse/testing/testing_zspmv.cpp
-index 2eff03f63f2a..776920a9cd4c 100644
--- a/sparse/testing/testing_zspmv.cpp
-+++ b/sparse/testing/testing_zspmv.cpp
-@@ -72,15 +72,6 @@
- 
- #define PRECISION_z
- 
-/* For hipSPARSE, they use a separate complex type than for hipBLAS */
-#if defined(MAGMA_HAVE_HIP)
-  #ifdef PRECISION_z
-    #define hipblasDoubleComplex hipDoubleComplex
-#elif defined(PRECISION_c)
-    #define hipblasComplex hipComplex
-  #endif
-#endif
-
- /* ////////////////////////////////////////////////////////////////////////////
-    -- testing sparse matrix vector product
- */
-diff --git a/src/shpotrf_gpu.cpp b/src/shpotrf_gpu.cpp
-index 6c57f2b57ab2..d981760d1057 100644
--- a/src/shpotrf_gpu.cpp
-+++ b/src/shpotrf_gpu.cpp
-@@ -55,10 +55,10 @@ magma_sgemm_fp16(
-     hipblasGemmEx( queue->hipblas_handle(),
- 		           hipblas_trans_const( transA ), hipblas_trans_const( transB ),
- 		           int(m), int(n), int(k),
-		           (void*)&alpha, (void*)dhA, HIPBLAS_R_16F, (int)lddha,
-                                  (void*)dhB, HIPBLAS_R_16F, (int)lddhb,
-		           (void*)&beta,  (void*)dC,  HIPBLAS_R_32F, (int)lddc,
-		           HIPBLAS_R_32F, HIPBLAS_GEMM_DEFAULT);
-+		           (void*)&alpha, (void*)dhA, HIP_R_16F, (int)lddha,
-+                                  (void*)dhB, HIP_R_16F, (int)lddhb,
-+		           (void*)&beta,  (void*)dC,  HIP_R_32F, (int)lddc,
-+		           HIPBLAS_COMPUTE_32F, HIPBLAS_GEMM_DEFAULT);
-     return 0;
-     #else
-     return MAGMA_ERR_NOT_SUPPORTED;
-diff --git a/src/xhsgetrf_gpu.cpp b/src/xhsgetrf_gpu.cpp
-index 0a0a0c7aee04..7b1e9d61586a 100644
--- a/src/xhsgetrf_gpu.cpp
-+++ b/src/xhsgetrf_gpu.cpp
-@@ -403,10 +403,10 @@ magma_xhsgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(nextjb), int(m-nextj), int(jb),
-                            &c_neg_one, dAT_hp(j,     nextj), HIPBLAS_R_16F, int(lddat),
-                                        dAT_hp(nextj, j    ), HIPBLAS_R_16F, int(lddat),
-                            &c_one,     dAT_hp(nextj, nextj), HIPBLAS_R_16F, int(lddat),
-                            HIPBLAS_R_32F, ALGO);
-+                            &c_neg_one, dAT_hp(j,     nextj), HIP_R_16F, int(lddat),
-+                                        dAT_hp(nextj, j    ), HIP_R_16F, int(lddat),
-+                            &c_one,     dAT_hp(nextj, nextj), HIP_R_16F, int(lddat),
-+                            HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_GEMEX_I16_O16_C16 ) {
-@@ -422,10 +422,10 @@ magma_xhsgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(nextjb), int(m-nextj), int(jb),
-                            &h_neg_one, dAT_hp(j,     nextj), HIPBLAS_R_16F, int(lddat),
-                                        dAT_hp(nextj, j    ), HIPBLAS_R_16F, int(lddat),
-                            &h_one,     dAT_hp(nextj, nextj), HIPBLAS_R_16F, int(lddat),
-                            HIPBLAS_R_16F, ALGO);
-+                            &h_neg_one, dAT_hp(j,     nextj), HIP_R_16F, int(lddat),
-+                                        dAT_hp(nextj, j    ), HIP_R_16F, int(lddat),
-+                            &h_one,     dAT_hp(nextj, nextj), HIP_R_16F, int(lddat),
-+                            HIPBLAS_COMPUTE_16F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_HGEMM ) {
-@@ -480,10 +480,10 @@ magma_xhsgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(maxn-(nextj+nextjb)), int(m-nextj), int(jb),
-                            &c_neg_one, dAT_hp(j,     nextj+nextjb), HIPBLAS_R_16F, int(lddat),
-                                        dAT_hp(nextj, j           ), HIPBLAS_R_16F, int(lddat),
-                            &c_one,     dAT_hp(nextj, nextj+nextjb), HIPBLAS_R_16F, int(lddat),
-                            HIPBLAS_R_32F, ALGO);
-+                            &c_neg_one, dAT_hp(j,     nextj+nextjb), HIP_R_16F, int(lddat),
-+                                        dAT_hp(nextj, j           ), HIP_R_16F, int(lddat),
-+                            &c_one,     dAT_hp(nextj, nextj+nextjb), HIP_R_16F, int(lddat),
-+                            HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_GEMEX_I16_O16_C16 ) {
-@@ -499,10 +499,10 @@ magma_xhsgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(maxn-(nextj+nextjb)), int(m-nextj), int(jb),
-                            &h_neg_one, dAT_hp(j,     nextj+nextjb), HIPBLAS_R_16F, int(lddat),
-                                        dAT_hp(nextj, j           ), HIPBLAS_R_16F, int(lddat),
-                            &h_one,     dAT_hp(nextj, nextj+nextjb), HIPBLAS_R_16F, int(lddat),
-                            HIPBLAS_R_16F, ALGO);
-+                            &h_neg_one, dAT_hp(j,     nextj+nextjb), HIP_R_16F, int(lddat),
-+                                        dAT_hp(nextj, j           ), HIP_R_16F, int(lddat),
-+                            &h_one,     dAT_hp(nextj, nextj+nextjb), HIP_R_16F, int(lddat),
-+                            HIPBLAS_COMPUTE_16F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_HGEMM ) {
-diff --git a/src/xshgetrf_gpu.cpp b/src/xshgetrf_gpu.cpp
-index 96ec1d97ffb8..345aefa1ad76 100644
--- a/src/xshgetrf_gpu.cpp
-+++ b/src/xshgetrf_gpu.cpp
-@@ -270,10 +270,10 @@ magma_xshgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
- 		                    hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
- 		                    int(nextjb), int(m-nextj), int(jb),
-		                    &c_neg_one, dAtrsm1_hp,        HIPBLAS_R_16F, int(maxnb),
-		                                dApanel_hp,        HIPBLAS_R_16F, int(jb),
-		                    &c_one,     dAT(nextj, nextj), HIPBLAS_R_32F, int(lddat),
-		                    HIPBLAS_R_32F, ALGO);
-+		                    &c_neg_one, dAtrsm1_hp,        HIP_R_16F, int(maxnb),
-+		                                dApanel_hp,        HIP_R_16F, int(jb),
-+		                    &c_one,     dAT(nextj, nextj), HIP_R_32F, int(lddat),
-+		                    HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_GEMEX_I32_O32_C32 ) {
-@@ -289,10 +289,10 @@ magma_xshgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(nextjb), int(m-nextj), int(jb),
-                            &c_neg_one, dAT(j,     nextj), HIPBLAS_R_32F, int(lddat),
-                                        dAT(nextj,     j), HIPBLAS_R_32F, int(lddat),
-                            &c_one,     dAT(nextj, nextj), HIPBLAS_R_32F, int(lddat),
-                            HIPBLAS_R_32F, ALGO);
-+                            &c_neg_one, dAT(j,     nextj), HIP_R_32F, int(lddat),
-+                                        dAT(nextj,     j), HIP_R_32F, int(lddat),
-+                            &c_one,     dAT(nextj, nextj), HIP_R_32F, int(lddat),
-+                            HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_SGEMM ) {
-@@ -343,10 +343,10 @@ magma_xshgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(maxn-(nextj+nextjb)), int(m-nextj), int(jb),
-                            &c_neg_one, dAtrsm2_hp              , HIPBLAS_R_16F, int(maxm),
-                                        dApanel_hp              , HIPBLAS_R_16F, int(jb),
-                            &c_one,     dAT(nextj, nextj+nextjb), HIPBLAS_R_32F, int(lddat),
-                            HIPBLAS_R_32F, ALGO);
-+                            &c_neg_one, dAtrsm2_hp              , HIP_R_16F, int(maxm),
-+                                        dApanel_hp              , HIP_R_16F, int(jb),
-+                            &c_one,     dAT(nextj, nextj+nextjb), HIP_R_32F, int(lddat),
-+                            HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_GEMEX_I32_O32_C32 ) {
-@@ -362,10 +362,10 @@ magma_xshgetrf_gpu(
-                     hipblasGemmEx( queues[1]->hipblas_handle(),
-                             hipblas_trans_const( MagmaNoTrans ), hipblas_trans_const( MagmaNoTrans ),
-                             int(maxn-(nextj+nextjb)), int(m-nextj), int(jb),
-                            &c_neg_one, dAT(j    , nextj+nextjb), HIPBLAS_R_32F, int(lddat),
-                                        dAT(nextj, j           ), HIPBLAS_R_32F, int(lddat),
-                            &c_one,     dAT(nextj, nextj+nextjb), HIPBLAS_R_32F, int(lddat),
-                            HIPBLAS_R_32F, ALGO);
-+                            &c_neg_one, dAT(j    , nextj+nextjb), HIP_R_32F, int(lddat),
-+                                        dAT(nextj, j           ), HIP_R_32F, int(lddat),
-+                            &c_one,     dAT(nextj, nextj+nextjb), HIP_R_32F, int(lddat),
-+                            HIPBLAS_COMPUTE_32F, ALGO);
-                     #endif
-                 }
-                 else if( mp_algo_type == Magma_MP_SGEMM ) {
-diff --git a/testing/testing_sgemm_fp16.cpp b/testing/testing_sgemm_fp16.cpp
-index 84f24b95a4a4..0fbfc0b338fe 100644
--- a/testing/testing_sgemm_fp16.cpp
-+++ b/testing/testing_sgemm_fp16.cpp
-@@ -59,10 +59,10 @@ magma_sgemm_fp16_v1(
-     hipblasGemmEx( magma_queue_get_hipblas_handle( queue ),
- 		           hipblas_trans_const( transA ), hipblas_trans_const( transB ),
- 		           int(m), int(n), int(k),
-		           (void*)&alpha, (void*)dhA, HIPBLAS_R_16F, (int)ldda,
-                                  (void*)dhB, HIPBLAS_R_16F, (int)lddb,
-		           (void*)&beta,  (void*)dC,  HIPBLAS_R_32F, (int)lddc,
-		           HIPBLAS_R_32F, HIPBLAS_GEMM_DEFAULT);
-+		           (void*)&alpha, (void*)dhA, HIP_R_16F, (int)ldda,
-+                                  (void*)dhB, HIP_R_16F, (int)lddb,
-+		           (void*)&beta,  (void*)dC,  HIP_R_32F, (int)lddc,
-+		           HIPBLAS_COMPUTE_32F, HIPBLAS_GEMM_DEFAULT);
-     #endif
-     return 0;
- }
-@@ -97,10 +97,10 @@ magma_sgemm_fp16_v2(
-     hipblasGemmEx( magma_queue_get_hipblas_handle( queue ),
- 		           hipblas_trans_const( transA ), hipblas_trans_const( transB ),
- 		           int(m), int(n), int(k),
-		           (void*)&alpha, (void*)dhA, HIPBLAS_R_16F, (int)ldda,
-                                  (void*)dhB, HIPBLAS_R_16F, (int)lddb,
-		           (void*)&beta,  (void*)dC,  HIPBLAS_R_32F, (int)lddc,
-		           HIPBLAS_R_32F, HIPBLAS_GEMM_DEFAULT);
-+		           (void*)&alpha, (void*)dhA, HIP_R_16F, (int)ldda,
-+                                  (void*)dhB, HIP_R_16F, (int)lddb,
-+		           (void*)&beta,  (void*)dC,  HIP_R_32F, (int)lddc,
-+		           HIPBLAS_COMPUTE_32F, HIPBLAS_GEMM_DEFAULT);
-     #endif
-     return 0;
- }
-diff --git a/testing/testing_zgemm_batched.cpp b/testing/testing_zgemm_batched.cpp
-index 0b300d1081be..350d0049865d 100644
--- a/testing/testing_zgemm_batched.cpp
-+++ b/testing/testing_zgemm_batched.cpp
-@@ -190,11 +190,11 @@ int main( int argc, char** argv)
-                   hipblasZgemmBatched(
-                                    opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB),
-                                    int(M), int(N), int(K),
-                                   (const hipblasDoubleComplex*)&alpha,
-                                   (const hipblasDoubleComplex**) d_A_array, int(ldda),
-                                   (const hipblasDoubleComplex**) d_B_array, int(lddb),
-                                   (const hipblasDoubleComplex*)&beta,
-                                   (hipblasDoubleComplex**)d_C_array, int(lddc), int(batchCount) );
-+                                   (const hipDoubleComplex*)&alpha,
-+                                   (const hipDoubleComplex**) d_A_array, int(ldda),
-+                                   (const hipDoubleComplex**) d_B_array, int(lddb),
-+                                   (const hipDoubleComplex*)&beta,
-+                                   (hipDoubleComplex**)d_C_array, int(lddc), int(batchCount) );
-                 #endif
-             }
-             else{
-@@ -211,11 +211,11 @@ int main( int argc, char** argv)
-                 hipblasZgemmStridedBatched(
-                                    opts.handle, cublas_trans_const(opts.transA), cublas_trans_const(opts.transB),
-                                    int(M), int(N), int(K),
-                                   (const hipblasDoubleComplex*)&alpha,
-                                   (const hipblasDoubleComplex*) d_A, int(ldda), ldda * An,
-                                   (const hipblasDoubleComplex*) d_B, int(lddb), lddb * Bn,
-                                   (const hipblasDoubleComplex*)&beta,
-                                   (hipblasDoubleComplex*)d_C, int(lddc), lddc*N, int(batchCount) );
-+                                   (const hipDoubleComplex*)&alpha,
-+                                   (const hipDoubleComplex*) d_A, int(ldda), ldda * An,
-+                                   (const hipDoubleComplex*) d_B, int(lddb), lddb * Bn,
-+                                   (const hipDoubleComplex*)&beta,
-+                                   (hipDoubleComplex*)d_C, int(lddc), lddc*N, int(batchCount) );
-                 #endif
-             }
- 
-diff --git a/testing/testing_zgemv_batched.cpp b/testing/testing_zgemv_batched.cpp
-index 9c63a405be17..be1dea4cfa6b 100644
--- a/testing/testing_zgemv_batched.cpp
-+++ b/testing/testing_zgemv_batched.cpp
-@@ -182,11 +182,11 @@ int main( int argc, char** argv)
-                 #else
-                 hipblasZgemvBatched(opts.handle, hipblas_trans_const(opts.transA),
-                                       M, N,
-                                      (const hipblasDoubleComplex *)&alpha,
-                                      (const hipblasDoubleComplex **)d_A_array, ldda,
-                                      (const hipblasDoubleComplex **)d_X_array, incx,
-                                      (const hipblasDoubleComplex *)&beta,
-                                      (hipblasDoubleComplex **)d_Y_array, incy, batchCount);
-+                                      (const hipDoubleComplex *)&alpha,
-+                                      (const hipDoubleComplex **)d_A_array, ldda,
-+                                      (const hipDoubleComplex **)d_X_array, incx,
-+                                      (const hipDoubleComplex *)&beta,
-+                                      (hipDoubleComplex **)d_Y_array, incy, batchCount);
-                 #endif
-             }
-             else{
-@@ -210,11 +210,11 @@ int main( int argc, char** argv)
-                 #else
-                 hipblasZgemvStridedBatched(opts.handle, hipblas_trans_const(opts.transA),
-                                       M, N,
-                                      (const hipblasDoubleComplex *)&alpha,
-                                      (const hipblasDoubleComplex *)d_A, ldda, ldda*N,
-                                      (const hipblasDoubleComplex *)d_X, incx, incx*Xm,
-                                      (const hipblasDoubleComplex *)&beta,
-                                      (hipblasDoubleComplex *)d_Y, incy, incy*Ym, batchCount);
-+                                      (const hipDoubleComplex *)&alpha,
-+                                      (const hipDoubleComplex *)d_A, ldda, ldda*N,
-+                                      (const hipDoubleComplex *)d_X, incx, incx*Xm,
-+                                      (const hipDoubleComplex *)&beta,
-+                                      (hipDoubleComplex *)d_Y, incy, incy*Ym, batchCount);
-                 #endif
-             }
-             device_time = magma_sync_wtime( opts.queue ) - device_time;
-diff --git a/testing/testing_zgeqrf_batched.cpp b/testing/testing_zgeqrf_batched.cpp
-index 6cc62ab63055..acdb92754f44 100644
--- a/testing/testing_zgeqrf_batched.cpp
-+++ b/testing/testing_zgeqrf_batched.cpp
-@@ -195,8 +195,8 @@ int main( int argc, char** argv)
-             #endif
-             #else
-             hipblasZgeqrfBatched( opts.handle, int(M), int(N),
-                                 (hipblasDoubleComplex**)dA_array, int(ldda),
-                                 (hipblasDoubleComplex**)dtau_array,
-+                                 (hipDoubleComplex**)dA_array, int(ldda),
-+                                 (hipDoubleComplex**)dtau_array,
-                                  &device_info, int(batchCount) );
-             #endif
- 
-diff --git a/testing/testing_zgetrf_batched.cpp b/testing/testing_zgetrf_batched.cpp
-index ea9871364901..d23de5d934bd 100644
--- a/testing/testing_zgetrf_batched.cpp
-+++ b/testing/testing_zgetrf_batched.cpp
-@@ -238,7 +238,7 @@ int main( int argc, char** argv)
-                                      dinfo_device, int(batchCount) );
-                 #else
-                 hipblasZgetrfBatched( opts.handle, int(N),
-                                     (hipblasDoubleComplex**)dA_array, int(ldda), dipiv_device,
-+                                     (hipDoubleComplex**)dA_array, int(ldda), dipiv_device,
-                                      dinfo_device, int(batchCount) );
-                 #endif
-             }
-diff --git a/testing/testing_ztrsm_batched.cpp b/testing/testing_ztrsm_batched.cpp
-index c2b6d0d88660..7a9f7c449dda 100644
--- a/testing/testing_ztrsm_batched.cpp
-+++ b/testing/testing_ztrsm_batched.cpp
-@@ -219,9 +219,9 @@ int main( int argc, char** argv)
-             hipblasZtrsmBatched(
-                 opts.handle, cublas_side_const(opts.side), cublas_uplo_const(opts.uplo),
-                 cublas_trans_const(opts.transA), cublas_diag_const(opts.diag),
-                int(M), int(N), (const hipblasDoubleComplex*)&alpha,
-                (hipblasDoubleComplex* const*) d_A_array, int(ldda),
-                (      hipblasDoubleComplex**) d_B_array, int(lddb), int(batchCount) );
-+                int(M), int(N), (const hipDoubleComplex*)&alpha,
-+                (hipDoubleComplex* const*) d_A_array, int(ldda),
-+                (      hipDoubleComplex**) d_B_array, int(lddb), int(batchCount) );
-             #endif
- 
-             cublas_time = magma_sync_wtime( opts.queue ) - cublas_time;
-diff --git a/tools/hipify-perl b/tools/hipify-perl
-index 1a7f889eb1eb..aa6a06ec5592 100755
--- a/tools/hipify-perl
-+++ b/tools/hipify-perl
-@@ -986,15 +986,8 @@ sub simpleSubstitutions {
-     $ft{'type'} += s/\bcsrilu02Info_t\b/csrilu02Info_t/g;
-     $ft{'type'} += s/\bcsrsv2Info_t\b/csrsv2Info_t/g;
-     $ft{'type'} += s/\bcuComplex\b/hipComplex/g;
-    
-    # original type subs
-    #$ft{'type'} += s/\bcuDoubleComplex\b/hipDoubleComplex/g;
-    #$ft{'type'} += s/\bcuFloatComplex\b/hipFloatComplex/g;
-    
-    # replace them with hipBLAS types
-    $ft{'type'} += s/\bcuDoubleComplex\b/hipblasDoubleComplex/g;
-    $ft{'type'} += s/\bcuFloatComplex\b/hipblasFloatComplex/g;
-    
-+    $ft{'type'} += s/\bcuDoubleComplex\b/hipDoubleComplex/g;
-+    $ft{'type'} += s/\bcuFloatComplex\b/hipFloatComplex/g;
-     $ft{'type'} += s/\bcublasDataType_t\b/hipblasDatatype_t/g;
-     $ft{'type'} += s/\bcublasDiagType_t\b/hipblasDiagType_t/g;
-     $ft{'type'} += s/\bcublasFillMode_t\b/hipblasFillMode_t/g;
-@@ -1210,9 +1203,9 @@ sub simpleSubstitutions {
-     $ft{'numeric_literal'} += s/\bCUDA_ERROR_UNKNOWN\b/hipErrorUnknown/g;
-     $ft{'numeric_literal'} += s/\bCUDA_ERROR_UNMAP_FAILED\b/hipErrorUnmapFailed/g;
-     $ft{'numeric_literal'} += s/\bCUDA_ERROR_UNSUPPORTED_LIMIT\b/hipErrorUnsupportedLimit/g;
-    $ft{'numeric_literal'} += s/\bCUDA_R_16F\b/HIPBLAS_R_16F/g;
-    $ft{'numeric_literal'} += s/\bCUDA_R_32F\b/HIPBLAS_R_32F/g;
-    $ft{'numeric_literal'} += s/\bCUDA_R_64F\b/HIPBLAS_R_64F/g;
-+    $ft{'numeric_literal'} += s/\bCUDA_R_16F\b/HIP_R_16F/g;
-+    $ft{'numeric_literal'} += s/\bCUDA_R_32F\b/HIP_R_32F/g;
-+    $ft{'numeric_literal'} += s/\bCUDA_R_64F\b/HIP_R_64F/g;
-     $ft{'numeric_literal'} += s/\bCUDA_SUCCESS\b/hipSuccess/g;
-     $ft{'numeric_literal'} += s/\bCUDNN_16BIT_INDICES\b/HIPDNN_16BIT_INDICES/g;
-     $ft{'numeric_literal'} += s/\bCUDNN_32BIT_INDICES\b/HIPDNN_32BIT_INDICES/g;
-diff --git a/tools/magmasubs.py b/tools/magmasubs.py
-index f1e0dabca776..00b8aed30295 100644
--- a/tools/magmasubs.py
-+++ b/tools/magmasubs.py
-@@ -501,8 +501,7 @@ subs = {
-     ('float',                'double',               'cuFloatComplex',       'cuDoubleComplex'     ),
-     ('float',                'double',               'hipFloatComplex',      'hipDoubleComplex'    ),
-     ('CUDA_R_32F',           'CUDA_R_64F',           'CUDA_C_32F',           'CUDA_C_64F'          ),
-    #('float',                'double',               'hipComplex',           'hipDoubleComplex'   ),
-    ('float',                'double',               'hipblasComplex',       'hipblasDoubleComplex'),
-+    ('float',                'double',               'hipComplex',           'hipDoubleComplex'    ),
-     ('float',                'double',               'MKL_Complex8',         'MKL_Complex16'       ),
-     ('magmaFloat_const_ptr', 'magmaDouble_const_ptr','magmaFloatComplex_const_ptr', 'magmaDoubleComplex_const_ptr'),  # before magmaDoubleComplex
-     ('magmaFloat_const_ptr', 'magmaDouble_const_ptr','magmaFloat_const_ptr',        'magmaDouble_const_ptr'       ),  # before magmaDoubleComplex
-- 
-2.51.0
-
@@ -1,100 +1,50 @@
-%bcond gitcommit 1
-%if %{with gitcommit}
-%global commit0 4e4e95f2e5d52a7237cd04a1aec4501463156396
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20260210
-%endif
+# 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 rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//' -e 's/-flto=thin//' )
-# lto problems
-# ld.lld: error: CMakeFiles/magma.dir/control/libmagma.so.2.9.0.lto.constants.cpp.o:(.rodata.magma_diag_const+0x4): relocation R_X86_64_PC32 out of range: 5412144436 is not in [-2147483648, 2147483647]; r
-%global _lto_cflags %nil
+%global toolchain clang

-# Need to have a GPU in the build machine to test.
-# To speed up testing, only gfx1201 is built
 %bcond test 0
-%if %{with test}
-%global gpu_list gfx1201
-%else
-%global gpu_list %{rocm_gpu_list_default}
-%endif
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif

 Name:           magma
-Version:        2.9.0
+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/
-License:        BSD-3-Clause AND MIT
-# From the license check
-# The main license is BSD-3-Clause
-#   COPYRIGHT
-# Then ICS, with file copied from OPEN BSD
-#   control/strlcpy.cpp
-# Then MIT, with a file copied from hipify
-# hipify is used but not delivered
-#   tools/hipify-perl
-# Reported GPL but not used, other similar files for cuda
-#   results/v1.5.0/cuda7.0-k40c/setup.txt
-
-# pickup newer commit for including Pull #27
-# https://github.com/icl-utk-edu/magma/pull/27
-%if %{with gitcommit}
-Source0:        https://github.com/icl-utk-edu/%{name}/archive/%{commit0}.tar.gz
-%else
+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
-%endif

-# For versioning the *.so's
-# https://bitbucket.org/icl/magma/issues/77/versioning-so
-Patch0:         0001-Prepare-magma-cmake-for-fedora.patch
+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

-# https://github.com/jeffdaily/magma/commit/1b966b72402e3f37ebd462f3d7e019e669e510ff
-Patch1:         0001-magma-ROCm-7-changes.patch
-
-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 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:  openblas-devel
-BuildRequires:  hipblas-devel
-BuildRequires:  hipsparse-devel
+BuildRequires:  hipcc
+BuildRequires:  lld
+BuildRequires:  llvm
 BuildRequires:  ninja
 BuildRequires:  python3
 BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
 BuildRequires:  rocm-llvm-macros
-BuildRequires:  rocm-comgr-devel
-# BuildRequires:  rocm-core-devel
-BuildRequires:  rocm-hip-devel
-BuildRequires:  rocr-runtime-devel
-#BuildRequires:  rocm-rpm-macros
-
-# MIT
-# Just the hipify-perl file is taken and it is very old
-# This is older than any release of https://github.com/ROCm/HIPIFY.git
-# So setting to earliest release -1
-Provides:       bundled(hipify) = 3.4.0
-
-ExclusiveArch:  x86_64 riscv64

 %description
 Matrix Algebra on GPU and Multi-core Architectures (MAGMA) is a collection
@@ -117,29 +67,22 @@ 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
+%package        devel
 Summary:        Libraries and headers for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description devel
+%description    devel
 %{summary}

-%prep
-%if %{with gitcommit}
-%setup -q -n %{name}-%{commit0}
-%else
-%autosetup -p1
-%endif
+%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}
-# Just the test gpu gfx1201
-sed -i -e 's@1032 1033@1201@' Makefile
-# Remove some tests,
+# Remove a test that fails to link (undefined magma_generate_matrix)
 sed -i -e '/testing_zgenerate.cpp/d' testing/Makefile.src
 %else
-# Add some more gfx's
-# https://bitbucket.org/icl/magma/issues/76/a-few-new-rocm-gpus
-sed -i -e 's@1032 1033@950 1032 1033 1100 1101 1102 1103 1150 1151 1152 1153 1200 1201@' Makefile
 # 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
@@ -152,9 +95,6 @@ sed -i -e 's@add_custom_target( sparse-testing DEPENDS ${sparse-testing} )@#add_
 sed -i -e 's@DESTINATION lib@DESTINATION ${CMAKE_INSTALL_LIBDIR}@' CMakeLists.txt
 sed -i -e 's@DESTINATION bin@DESTINATION ${CMAKE_INSTALL_BINDIR}@' CMakeLists.txt

-# Version *.so
-sed -i -e 's@magma_VERSION@"%{version}"@g' 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
@@ -169,67 +109,14 @@ sed -i -e '/strlcpy/d' include/magma_auxiliary.h
 sed -i -e 's@magma_strlcpy@strlcpy@' control/trace.cpp
 rm control/strlcpy.cpp

-# Policy CMP0037 may not be set to OLD behavior because this version of CMake
-sed -i -e 's@cmake_policy( SET CMP0037 OLD)@#cmake_policy( SET CMP0037 OLD)@' CMakeLists.txt
-
-# Add offload-compress compile flags
-sed -i -e 's@-DROCM_VERSION@--offload-compress -DROCM_VERSION@' CMakeLists.txt
-
-%build
-
-#export HIP_PATH=`hipconfig -p`
-#export ROCM_PATH=`hipconfig -R`
-
+%build -p
 echo "BACKEND = hip"                          > make.inc
 echo "FORT = false"                          >> make.inc
-%if %{with test}
-echo "GPU_TARGET = gfx1201" >> make.inc
-%else
-echo "GPU_TARGET = gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx942;gfx950;gfx1010;gfx1012;gfx1030;gfx1031;gfx1035;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201" >> make.inc
-%endif
+echo "GPU_TARGET = gfx1100;gfx1200;gfx1201"  >> make.inc

 make generate
-
-%cmake -G Ninja \
-       -DBLA_VENDOR=OpenBLAS \
-       -DCMAKE_BUILD_TYPE=%{build_type} \
-       -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-       -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-       -DAMDGPU_TARGETS=%{gpu_list} \
-       -DCMAKE_INSTALL_LIBDIR=%_libdir \
-       -DMAGMA_ENABLE_HIP=ON \
-       -DUSE_FORTRAN=OFF
-
-%cmake_build
-
-%install
-%cmake_install
-
 %if %{with test}
 %check
-# Results should be something like
-# % MAGMA 2.9.0 svn 32-bit magma_int_t, 64-bit pointer.
-# % HIP runtime 70051831, driver 70051831. MAGMA not compiled with OpenMP. 
-# % device 0: AMD Radeon Graphics, 2420.0 MHz clock, 16304.0 MiB memory, gcn arch gfx1201
-# % Sat Oct  4 08:05:46 2025
-# % Usage: redhat-linux-build/testing/testing_sgemm [options] [-h|--help]
-#
-# % If running lapack (option --lapack), MAGMA and HIPBLAS error are both computed
-# % relative to CPU BLAS result. Else, MAGMA error is computed relative to HIPBLAS result.
-#
-# % transA = No transpose, transB = No transpose
-# %   M     N     K   MAGMA Gflop/s (ms)  HIPBLAS Gflop/s (ms)   CPU Gflop/s (ms)  MAGMA error  HIPBLAS error
-# %========================================================================================================
-# 1088  1088  1088   1778.70 (   1.45)     146.15 (  17.63)     ---   (  ---  )    1.25e-08        ---    ok
-# 2112  2112  2112   10940.92 (   1.72)    10784.16 (   1.75)     ---   (  ---  )    1.19e-08        ---    ok
-# 3136  3136  3136   10919.38 (   5.65)    11215.23 (   5.50)     ---   (  ---  )    1.34e-08        ---    ok
-# 4160  4160  4160   11144.04 (  12.92)    12124.94 (  11.87)     ---   (  ---  )    1.13e-08        ---    ok
-# 5184  5184  5184   12999.31 (  21.43)    14869.66 (  18.74)     ---   (  ---  )    1.29e-08        ---    ok
-# 6208  6208  6208   13411.76 (  35.68)    14849.32 (  32.22)     ---   (  ---  )    1.07e-08        ---    ok
-# 7232  7232  7232   13335.70 (  56.73)    14755.09 (  51.27)     ---   (  ---  )    8.81e-09        ---    ok
-# 8256  8256  8256   13441.97 (  83.73)    14693.99 (  76.59)     ---   (  ---  )    8.36e-09        ---    ok
-# 9280  9280  9280   13347.32 ( 119.75)    14706.62 ( 108.68)     ---   (  ---  )    1.13e-08        ---    ok
-# 10304 10304 10304   13291.48 ( 164.62)    14655.85 ( 149.29)     ---   (  ---  )    1.04e-08        ---    ok
 %{_vpath_builddir}/testing/testing_sgemm
 %endif

@@ -245,5 +132,4 @@ make generate
 %{_libdir}/libmagma_sparse.so

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 2.9.0-1
- Import from upstream
+%autochangelog
@@ -1,50 +0,0 @@
-From a8a1e5626512209fce844a273f55ef4b708c9249 Mon Sep 17 00:00:00 2001
-From: Tom Rix <Tom.Rix@amd.com>
-Date: Sun, 21 Sep 2025 08:24:27 -0700
-Subject: [PATCH] miopen add link and compile pools
-
---
- CMakeLists.txt | 27 +++++++++++++++++++++++++++
- 1 file changed, 27 insertions(+)
-
-diff --git a/CMakeLists.txt b/CMakeLists.txt
-index 514eb46ee71a..ea431a93ddc1 100644
--- a/CMakeLists.txt
-+++ b/CMakeLists.txt
-@@ -179,6 +179,33 @@ option(MIOPEN_EMBED_BUILD "Build with the set of embed flags." Off)
- option(MIOPEN_DISABLE_USERDB "Disable user database access" ${MIOPEN_EMBED_BUILD})
- option(MIOPEN_BUILD_CK "Build own CK libs inline with MIOpen build" OFF)
- 
-+#
-+# Seperate linking jobs from compiling
-+# Too many concurrent linking jobs can break the build
-+# Copied from LLVM
-+set(MIOPEN_PARALLEL_LINK_JOBS "" CACHE STRING
-+  "Define the maximum number of concurrent link jobs (Ninja only).")
-+if(CMAKE_GENERATOR MATCHES "Ninja")
-+  if(MIOPEN_PARALLEL_LINK_JOBS)
-+    set_property(GLOBAL APPEND PROPERTY JOB_POOLS link_job_pool=${MIOPEN_PARALLEL_LINK_JOBS})
-+    set(CMAKE_JOB_POOL_LINK link_job_pool)
-+  endif()
-+elseif(MIOPEN_PARALLEL_LINK_JOBS)
-+  message(WARNING "Job pooling is only available with Ninja generators.")
-+endif()
-+
-+set(MIOPEN_PARALLEL_COMPILE_JOBS "" CACHE STRING
-+  "Define the maximum number of concurrent compile jobs (Ninja only).")
-+if(CMAKE_GENERATOR MATCHES "Ninja")
-+  if(MIOPEN_PARALLEL_COMPILE_JOBS)
-+    set_property(GLOBAL APPEND PROPERTY JOB_POOLS compile_job_pool=${MIOPEN_PARALLEL_COMPILE_JOBS})
-+    set(CMAKE_JOB_POOL_COMPILE compile_job_pool)
-+  endif()
-+elseif(MIOPEN_PARALLEL_COMPILE_JOBS)
-+  message(WARNING "Job pooling is only available with Ninja generators.")
-+endif()
-+
-+
- # MIOPEN_USE_HIP_KERNELS is a Workaround for COMgr issues
- if(MIOPEN_EMBED_BUILD)
-     set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build as a shared library" FORCE)
-- 
-2.51.0
-
@@ -1,357 +1,147 @@
-%bcond gitcommit 0
-%if %{with gitcommit}
-%global commit0 2584e35062ad9c2edb68d93c464cf157bc57e3b0
-%global shortcommit0 %(c=%{commit0}; echo ${c:0:7})
-%global date0 20250926
-%endif
+# 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 upstreamname MIOpen
 %global rocm_release 7.1
-%global rocm_patch 1
+%global rocm_patch   1
 %global rocm_version %{rocm_release}.%{rocm_patch}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
+# rocm stack builds with clang
+%global toolchain clang

-%global miopen_name miopen%{pkg_suffix}
-
-%global toolchain rocm
-
-# hipcc does not support some clang flags
-# build_cxxflags does not honor CMAKE_BUILD_TYPE, strip out -g
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-g / /' -e 's/-mtls-dialect=gnu2//')
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# For testing
-# hardcoded use of gtest and dirs is not suitable for mock building
-# Testsuite is not in great shape, fails instead of skips ck tests
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%else
-%global build_test OFF
-%endif
-
-# Change this to the gpu family you are testing on
-%bcond check 0
-%global gpu_test default
-%if %{with test}
-%if %{with check}
-# Do not build everything to do the test on one thing
-%global rocm_gpu_list %{gpu_test}
-%endif
-%endif
-
-# Needs to match rocblas
-%bcond tensile 1
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-%global gpu_list %{rocm_gpu_list_default}
-%global _gpu_list gfx1100
-
-# Use ninja if it is available
-%bcond ninja 1
-
-%if %{with ninja}
-%global cmake_generator -G Ninja
-%else
-%global cmake_generator %{nil}
-%endif
-
-Name:           %{miopen_name}
-%if %{with gitcommit}
-Version:        git%{date0}.%{shortcommit0}
-Release:        1%{?dist}
-%else
+Name:           miopen
 Version:        %{rocm_version}
-Release:        1%{?dist}
-%endif
+Release:        %autorelease
 Summary:        AMD's Machine Intelligence Library
-License:        MIT AND BSD-2-Clause AND Apache-2.0 AND %{?fedora:LicenseRef-Fedora-Public-Domain}%{?suse_version:SUSE-Public-Domain}
-# The base license is MIT with a couple of exceptions
-# BSD-2-Clause
-#   driver/mloSoftmaxHost.hpp
-#   src/include/miopen/mlo_internal.hpp
-# Apache-2.0
-#   src/include/miopen/kernel_cache.hpp
-#   src/kernel_cache.cpp
-# Public Domain
-#   src/md5.cpp
+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

-%if %{with gitcommit}
-Url:            https://github.com/ROCm/rocm-libraries
-Source0:        %{url}/archive/%{commit0}/rocm-libraries-%{shortcommit0}.tar.gz
+# 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
-Url:            https://github.com/ROCm/%{upstreamname}
-Source0:        %{url}/archive/rocm-%{version}.tar.gz#/%{upstreamname}-%{version}.tar.gz
+BuildOption(conf):  -DBUILD_TESTING=OFF
 %endif
-
-# So we do not thrash memory
-Patch1:         0001-miopen-add-link-and-compile-pools.patch
-
-BuildRequires:  cmake
-BuildRequires:  pkgconfig(eigen3)
-BuildRequires:  gcc-c++
-BuildRequires:  fdupes
-
-BuildRequires:  llvm
-BuildRequires:  llvm-devel
-BuildRequires:  clang
-BuildRequires:  clang-devel
-BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
-BuildRequires:  fplus-devel
-BuildRequires:  frugally-deep-devel
-BuildRequires:  half-devel
-BuildRequires:  pkgconfig(libzstd)
-BuildRequires:  pkgconfig(nlohmann_json)
-BuildRequires:  hipblas%{pkg_suffix}-devel
-BuildRequires:  rocblas%{pkg_suffix}-devel
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocrand%{pkg_suffix}-devel
-BuildRequires:  roctracer%{pkg_suffix}-devel
-BuildRequires:  pkgconfig(sqlite3)
-BuildRequires:  zlib-devel
+# Disable optional backends not yet packaged on openRuyi
+BuildOption(conf):  -DMIOPEN_USE_COMPOSABLEKERNEL=OFF
+BuildOption(conf):  -DMIOPEN_USE_MLIR=OFF

 BuildRequires:  boost-devel
-BuildRequires:  pkgconfig(bzip2)
-
+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:  gmock-devel
-BuildRequires:  gtest-devel
+BuildRequires:  cmake(GTest)
 %endif
-
-%if %{with ninja}
+BuildRequires:  clang
+BuildRequires:  clang-tools-extra
+BuildRequires:  compiler-rt
+BuildRequires:  half
+BuildRequires:  hipcc
+BuildRequires:  hipblas-common-devel
+BuildRequires:  lld
+BuildRequires:  llvm
 BuildRequires:  ninja
-%define __builder ninja
-%endif
+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

-Provides:       miopen%{pkg_suffix} = %{version}-%{release}
-
-# Use ROCm devel at runtime
-Requires:       rocm-hip%{pkg_suffix}-devel
-Requires:       rocrand%{pkg_suffix}-devel
-# Also needs c++ to work
-# From pytorch-examples/word_language_model
-# MIOpen(HIP): Error [BuildHip] HIPRTC status = HIPRTC_ERROR_COMPILATION (6), source file: MIOpenDropoutHIP.cpp
-# MIOpen(HIP): Warning [BuildHip] In file included from /tmp/comgr-78a343/input/MIOpenDropoutHIP.cpp:32:
-# In file included from /tmp/comgr-78a343/include/miopen_rocrand.hpp:45:
-# In file included from /usr/include/rocrand/rocrand_xorwow.h:24:
-# /usr/include/rocrand/rocrand_common.h:39:10: fatal error: 'utility' file not found
-#   39 | #include <utility>
-#      |          ^~~~~~~~~
+Requires:       cmake(hip)
+Requires:       cmake(rocrand)
 Requires:       gcc-c++

-
-ExclusiveArch:  x86_64 riscv64
-
 %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}
+%package        devel
+Summary:        Libraries and headers for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Provides:       miopen%{pkg_suffix}-devel = %{version}-%{release}

-%description devel
+%description    devel
 %{summary}

 %if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
 %endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/miopen
-%patch -P1 -p1
-%else
-%autosetup -p1 -n %{upstreamname}-rocm-%{version}
-%endif
-
-# Readme has executable bit
-chmod 644 README.md
-
-# clang-tidy is brittle and not needed for rebuilding from a tarball
+%prep -a
+# clang-tidy is brittle and not needed when rebuilding from a tarball
 sed -i -e 's@clang-tidy@true@' cmake/ClangTidy.cmake

-# workaround error on finding lbunzip2
-sed -i -e 's@lbunzip2 bunzip2@bunzip2@' CMakeLists.txt
-
-# https://github.com/ROCm/MIOpen/issues/2672
-sed -i -e 's@find_path(HALF_INCLUDE_DIR half/half.hpp)@#find_path(HALF_INCLUDE_DIR half/half.hpp)@' CMakeLists.txt
-# #include <half/half.hpp> -> <half.hpp>
-for f in `find . -type f -name '*.hpp' -o -name '*.cpp' `; do
-    sed -i -e 's@#include <half/half.hpp>@#include <half.hpp>@' $f
-done
-# On 6.4.0
-# ../test/verify.hpp:198:56: error: no member named 'expr' in namespace 'half_float::detail'
-#  198 |     if constexpr(std::is_same_v<T, half_float::detail::expr>)
-# This is not our float, hack it out
+# 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

-# Tries to download its own googletest
-# No good knob to turn it off so hack the cmake
+# 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

-%if %{without tensile}
-sed -i -e 's@#define ROCBLAS_BETA_FEATURES_API 1@#define ROCBLAS_BETA_FEATURES_API 0@' src/include/miopen/handle.hpp
-sed -i -e 's@#define ROCBLAS_BETA_FEATURES_API 1@#define ROCBLAS_BETA_FEATURES_API 0@' src/solver/mha/mha_common.hpp
-sed -i -e 's@#define ROCBLAS_BETA_FEATURES_API 1@#define ROCBLAS_BETA_FEATURES_API 0@' src/gemm_v2.cpp
-%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

-# Our use of modules confuse install locations
-# The db is not installed relative to the lib dir.
-# Hardcode its location
-sed -i -e 's@GetLibPath().parent_path() / "share/miopen/db"@"/usr/share/miopen/db"@' src/db_path.cpp.in
-
-# Unsupported compiler flags
+# -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

-# Paths to clang
+# 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

-%build
-%if %{with gitcommit}
-cd projects/miopen
-%endif
-
-
-# Real cores, No hyperthreading
-COMPILE_JOBS=`cat /proc/cpuinfo | grep -m 1 'cpu cores' | awk '{ print $4 }'`
-if [ ${COMPILE_JOBS}x = x ]; then
-    COMPILE_JOBS=1
-fi
-# Take into account memmory usage per core, do not thrash real memory
-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
-LINK_MEM=32
-LINK_JOBS=`eval "expr 1 + ${MEM_GB} / ${LINK_MEM}"`
-
-#%{?suse_version:%{?build_cxxflags:CXXFLAGS="%{build_cxxflags}"}}
-
-%cmake %{cmake_generator} \
-       -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-       -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-       -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-       -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-       -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-       -DROCM_SYMLINK_LIBS=OFF \
-       -DHIP_PLATFORM=amd \
-       -DGPU_TARGETS=%{gpu_list} \
-       -DBUILD_TESTING=%{build_test} \
-       %{?build_type:-DCMAKE_BUILD_TYPE=%{build_type}} \
-       -DCMAKE_SKIP_RPATH=ON \
-       -DBoost_USE_STATIC_LIBS=OFF \
-       -DMIOPEN_PARALLEL_COMPILE_JOBS=$COMPILE_JOBS \
-       -DMIOPEN_PARALLEL_LINK_JOBS=$LINK_JOBS \
-       -DMIOPEN_BACKEND=HIP \
-       -DMIOPEN_BUILD_DRIVER=OFF \
-       -DMIOPEN_ENABLE_AI_IMMED_MODE_FALLBACK=OFF \
-       -DMIOPEN_ENABLE_AI_KERNEL_TUNING=OFF \
-       -DMIOPEN_TEST_ALL=%{build_test} \
-       -DMIOPEN_USE_HIPBLASLT=OFF \
-       -DMIOPEN_USE_MLIR=OFF \
-       -DMIOPEN_USE_COMPOSABLEKERNEL=OFF
-
-%cmake_build
-
-%if %{with test}
-%cmake_build -t tests
-%endif
-
-%if %{with test}
-%if %{with check}
-%check
-find . -name 'libMIOpen.so.1'
-export LD_LIBRARY_PATH=${PWD}/%{_vpath_builddir}/lib:$LD_LIBRARY_PATH
-%ctest
-%endif
-%endif
-
-%install
-%if %{with gitcommit}
-cd projects/miopen
-%endif
-%cmake_install
-
-# Extra license
-rm -f %{buildroot}%{pkg_prefix}/share/doc/miopen-hip/LICENSE.md
-
-%fdupes %{buildroot}%{pkg_prefix}
-
-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/miopen-hip/LICENSE.md

 %files
-%if %{with gitcommit}
-%doc projects/miopen/README.md
-%license projects/miopen/LICENSE.md
-%else
 %doc README.md
 %license LICENSE.md
-%endif
-
-%{pkg_prefix}/%{pkg_libdir}/libMIOpen.so.1{,.*}
-%{pkg_prefix}/libexec/miopen/
+%{_libdir}/libMIOpen.so.1{,.*}
+%{_libexecdir}/miopen/

 %files devel
-%{pkg_prefix}/share/miopen/
-%{pkg_prefix}/include/miopen/
-%{pkg_prefix}/%{pkg_libdir}/libMIOpen.so
-%{pkg_prefix}/%{pkg_libdir}/cmake/miopen/
+%{_datadir}/miopen/
+%{_includedir}/miopen/
+%{_libdir}/cmake/miopen/
+%{_libdir}/libMIOpen.so

 %if %{with test}
 %files test
-%{pkg_prefix}/bin/test*
+%{_bindir}/test*
 %endif

 %changelog
-* Mon Feb 2 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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
@@ -1,109 +1,60 @@
-%global upstreamname RCCL
+# 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}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}/
-%global pkg_suffix -%{rocm_release}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%endif
-%global rccl_name rccl%{pkg_suffix}
+# rocm stack builds with clang
+%global toolchain clang

-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-flto=thin//' -e 's/-mtls-dialect=gnu2//')
-
-%global _lto_cflags %{nil}
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# downloads tests, use mock --enable-network
-%bcond test 0
-%if %{with test}
-%global build_test ON
-%global __brp_check_rpaths %{nil}
-%else
-%global build_test OFF
-%endif
-
-%bcond export 0
-%if %{with export}
-%global build_compile_db ON
-%else
-%global build_compile_db OFF
-%endif
-
-# rccl is not supported on gfx1103
-# On 6.1.1
-# lld: error: ld-temp.o <inline asm>:1:25: specified hardware register is not supported on this GPU
-#        s_getreg_b32 s1, hwreg(HW_REG_HW_ID)
-#
-# On 6.2
-# Problems reported with gfx10, removing gfx10 and default (gfx10 and gfx11) from build list
-#
-%global gpu_list %{rocm_gpu_list_rccl}
-%global _gpu_list gfx1100
-
-Name:           %{rccl_name}
+Name:           rccl
 Version:        %{rocm_version}
-Release:        1%{?dist}
+Release:        %autorelease
 Summary:        ROCm Communication Collectives Library
-
-Url:            https://github.com/ROCm/rccl
 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

-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz#/%{upstreamname}-%{rocm_version}.tar.gz
+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:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  hipify%{pkg_suffix}
-BuildRequires:  gcc-c++
-BuildRequires:  fmt-devel
+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%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-core%{pkg_suffix}-devel
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocm-smi%{pkg_suffix}-devel
-
-%if %{with test}
-BuildRequires:  gtest-devel
-%endif
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-llvm-macros

 Requires:       %{name}-data = %{version}-%{release}
-Provides:       rccl%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64

 %description
 RCCL (pronounced "Rickle") is a stand-alone library of standard
@@ -122,144 +73,61 @@ 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.

-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
+%package        devel
 Summary:        Headers and libraries for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}
-Provides:       rccl%{pkg_suffix}-devel = %{version}-%{release}
+Provides:       rccl-devel = %{version}-%{release}

-%description devel
+%description    devel
 Headers and libraries for %{name}

-%package data
+%package        data
 Summary:        Data for %{name}
 BuildArch:      noarch

-%description data
+%description    data
 Data for %{name}

-%if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
-%endif
-
-%prep
-%autosetup -p1 -n rccl-rocm-%{version}
-
-# Allow user to set AMDGPU_TARGETS
-sed -i -e '/AMD GPU targets to compile for/d' CMakeLists.txt
-
-# No /opt/rocm/.info/version
-sed -i -e 's@cat ${ROCM_PATH}/.info/version@echo %{rocm_version}@' CMakeLists.txt
-
-# wrong path
-# https://github.com/ROCm/rccl/issues/1649
-sed -i -e 's@rocm-core/rocm_version.h@rocm_version.h@' src/include/hip_rocm_version_info.h
-
-# Problems building on SUSE
-# ENABLE_MSCCLPP=OFF
-sed -i -e 's@if (ENABLE_MSCCLPP AND NOT(${HOST_OS_ID} STREQUAL "ubuntu" OR ${HOST_OS_ID} STREQUAL "centos"))@if (ENABLE_MSCCLPP)@' CMakeLists.txt
-
-# Building --with test
-# .../test/common/TestBed.cpp:607:16: error: no member named 'setfill' in namespace 'std'
-#   607 |     ss << std::setfill(' ') << std::setw(20) << ncclFuncNames[funcType] << " ";
-# https://github.com/ROCm/rccl/issues/1749
-sed -i '/#include <map.*/a#include <iomanip>' test/common/TestBed.hpp
-
-# On Tumbleweed Q3,2025
-# /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)@' CMakeLists.txt

+%prep -a
 # Do not force install
 sed -i -e 's@set(CMAKE_INSTALL_LIBDIR@#set(CMAKE_INSTALL_LIBDIR@' cmake/Dependencies.cmake
-
-# RCCL uses -parallel-jobs for both compiling and linking
-# compiling is set to 12, which may be more than the cores on the build machine.
-# linking is set by reserving 16GB pre thread, can be too little.
-# Use our own heuristics here
-# Real cores, No hyperthreading
-COMPILE_JOBS=`cat /proc/cpuinfo | grep -m 1 'cpu cores' | awk '{ print $4 }'`
-if [ ${COMPILE_JOBS}x = x ]; then
-    COMPILE_JOBS=1
-fi
-# Try again..
-if [ ${COMPILE_JOBS} = 1 ]; then
-    COMPILE_JOBS=`lscpu | grep '^CPU(s)' | awk '{ print $2 }'`
-    if [ ${COMPILE_JOBS}x = x ]; then
-        COMPILE_JOBS=4
-    fi
-fi
-
-# Take into account memmory usage per core, do not thrash real memory
-# inflate this to prevent competing with normal compile jobs
-BUILD_MEM=16
-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
-LINK_MEM=65
-LINK_JOBS=`eval "expr 1 + ${MEM_GB} / ${LINK_MEM}"`
-
-sed -i -e "s@rccl PRIVATE -parallel-jobs=12@rccl PRIVATE -parallel-jobs=${COMPILE_JOBS}@" CMakeLists.txt
-sed -i -e "s@-parallel-jobs=\${num_linker_jobs}@-parallel-jobs=${LINK_JOBS}@" CMakeLists.txt
+# -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
-%cmake \
-    -DGPU_TARGETS=%{gpu_list} \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DBUILD_TESTS=%{build_test} \
-    -DCMAKE_BUILD_TYPE=%{build_type} \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_EXPORT_COMPILE_COMMANDS=%{build_compile_db} \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_SKIP_RPATH=ON \
-    -DENABLE_MSCCLPP=OFF \
-    -DEXPLICIT_ROCM_VERSION=%{rocm_version} \
-    -DHIP_PLATFORM=amd \
-    -DRCCL_ROCPROFILER_REGISTER=OFF \
-    -DROCM_PATH=%{pkg_prefix} \
-    -DROCM_SYMLINK_LIBS=OFF
-
+# 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
-%cmake_install
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/rccl/LICENSE.txt
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/rccl/LICENSE.txt

 %files
 %license LICENSE.txt
-%{pkg_prefix}/%{pkg_libdir}/librccl.so.*
-%{pkg_prefix}/bin/rcclras
+%{_libdir}/librccl.so.*
+%{_bindir}/rcclras

 %files data
-%{pkg_prefix}/share/rccl/msccl-algorithms/
-%{pkg_prefix}/share/rccl/msccl-unit-test-algorithms/
+%{_datadir}/rccl/msccl-algorithms/
+%{_datadir}/rccl/msccl-unit-test-algorithms/

 %files devel
 %doc README.md
-%{pkg_prefix}/include/rccl/
-%{pkg_prefix}/%{pkg_libdir}/cmake/rccl/
-%{pkg_prefix}/%{pkg_libdir}/librccl.so
+%{_includedir}/rccl/
+%{_libdir}/cmake/rccl/
+%{_libdir}/librccl.so

-%if %{with test}
 %files test
-%{pkg_prefix}/bin/rccl-UnitTests
-%endif
+%{_bindir}/rccl-UnitTests

 %changelog
-* Mon Jan 26 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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}
@@ -1,234 +1,111 @@
-%global upstreamname rocFFT
+# 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_patch 1 
 %global rocm_version %{rocm_release}.%{rocm_patch}

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%endif
-%global rocfft_name rocfft%{pkg_suffix}
+# rocm stack builds with clang
+%global toolchain clang

-%global toolchain rocm
-
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host-fcf-protection/')
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# kernel oops on gfx1201
-# https://github.com/ROCm/rocFFT/issues/560
-%bcond test 0
-%if %{with test}
-# Disable rpatch checks for a local build
-%global __brp_check_rpaths %{nil}
-%global build_test ON
-%else
-%global build_test OFF
-%endif
-
-# Option to test suite for testing on real HW:
-# May have to set gpu under test with
-# export HIP_VISIBLE_DEVICES=<num> - 0, 1 etc.
-%bcond check 0
-
-# For docs
-%bcond doc 0
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-# Use rocm-llvm strip
-%global __strip %rocmllvm_bindir/llvm-strip
-
-# Use ninja if it is available
-%bcond ninja 1
-
-%if %{with ninja}
-%global cmake_generator -G Ninja
-%else
-%global cmake_generator %{nil}
-%endif
-
-%global cmake_config \\\
-  -DBUILD_CLIENTS_TESTS_OPENMP=OFF \\\
-  -DBUILD_CLIENTS_TESTS=%{build_test} \\\
-  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \\\
-  -DCMAKE_BUILD_TYPE=%{build_type} \\\
-  -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \\\
-  -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \\\
-  -DCMAKE_CXX_FLAGS="--rtlib=compiler-rt --unwindlib=libgcc -fPIC" \\\
-  -DCMAKE_C_IMPLICIT_INCLUDE_DIRECTORIES=/usr/include \\\
-  -DCMAKE_CXX_IMPLICIT_INCLUDE_DIRECTORIES=/usr/include \\\
-  -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \\\
-  -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \\\
-  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \\\
-  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \\\
-  -DHIPCC_BIN_DIR=%{_bindir} \\\
-  -DHIP_COMPILER=%rocmllvm_bindir/clang++ \\\
-  -DHIP_PLATFORM=amd \\\
-  -DROCFFT_BUILD_OFFLINE_TUNER=OFF \\\
-  -DROCFFT_KERNEL_CACHE_ENABLE=OFF \\\
-  -DROCM_SYMLINK_LIBS=OFF \\\
-  -DSQLITE_USE_SYSTEM_PACKAGE=ON \\\
-  -Dhip_DIR=/usr/lib64/cmake/hip \\\
-  -DCMAKE_PREFIX_PATH="%{rocmllvm_cmakedir}/..;%{_libdir}/cmake"
-#   -DHIP_COMMON_DIR=$p/hip
-
-%global gpu_list %{rocm_gpu_list_default}
-%global _gpu_list gfx1100
-
-Name:           rocfft%{pkg_suffix}
+Name:           rocfft
 Version:        %{rocm_version}
-Release:        1%{?dist}
-Summary:        ROCm Fast Fourier Transforms (FFT) library
+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

-Url:            https://github.com/ROCm/%{upstreamname}
-Source0:        %{url}/archive/rocm-%{version}.tar.gz#/%{upstreamname}-rocm-%{version}.tar.gz
+Patch0:         0001-cmake-use-gnu-installdirs.patch

-BuildRequires:  llvm
-BuildRequires:  llvm-devel
-BuildRequires:  clang
-BuildRequires:  clang-devel
-BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
+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:  python3
-
-BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  pkgconfig(sqlite3)
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel >= %{rocm_release}
-
-%if %{with test}
-BuildRequires:  rocrand%{pkg_suffix}-devel
-BuildRequires:  fftw-devel
 BuildRequires:  boost-devel
-BuildRequires:  hiprand%{pkg_suffix}-devel
-
-BuildRequires:  gtest-devel
-
-# rocfft-test compiles some things and requires rocm-hip-devel
-Requires:  rocm-hip%{pkg_suffix}-devel >= %{rocm_release}
-
-%endif
-
-%if %{with doc}
-BuildRequires:  python3dist(sphinx)
-%endif
-
-%if %{with ninja}
+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
-%define __builder ninja
-%endif
-
-Provides:       rocfft%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64
-
-Patch0: 0001-cmake-use-gnu-installdirs.patch
+BuildRequires:  pkgconfig(fftw3)
+BuildRequires:  pkgconfig(sqlite3)
+BuildRequires:  python3
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros

 %description
-A library for computing Fast Fourier Transforms (FFT), part of ROCm.
+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
+%package        devel
 Summary:        The rocFFT development package
-Requires:       %{rocfft_name}%{?_isa} = %{version}-%{release}
-Requires:       rocm-hip%{pkg_suffix}-devel
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+Requires:       cmake(hip)

-%description devel
+%description    devel
 The rocFFT development package.

-%if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
-%endif
-
-%prep
-%autosetup -n %{upstreamname}-rocm-%{version} -p 1

+%prep -a
 # Do not care so much about the sqlite version
 sed -i -e 's@SQLite3 3.50.2 @SQLite3 @' cmake/sqlite.cmake

-%build
-# ensuring executables are PIE enabled
-export LDFLAGS="${LDFLAGS} -pie"
-
-# OpenMP tests are disabled because upstream sets rpath in that case without
-# a way to skip
-#
-# RHEL 9 has an issue with missing symbol __truncsfhf2 in libgcc.
-# So switch from libgcc to rocm-llvm's libclang-rt.builtins with
-# the rtlib=compiler-rt. Leave unwind unchange with unwindlib=libgcc
-%cmake %{cmake_generator} %{cmake_config} \
-    -DGPU_TARGETS=%{gpu_list}
-
-%cmake_build
-
-%install
-%cmake_install
-
-# we don't need the rocfft_rtc_helper binary, don't package it
+%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

-# we don't need or want the client-info file installed by rocfft
-rm -rf %{buildroot}/%{pkg_prefix}/.info
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/rocfft/LICENSE.md
-
-
-%check
 %if %{with test}
-%if %{with check}
+%check
 %{_vpath_builddir}/clients/staging/rocfft-test
 %endif
-%endif

-%files -n %{rocfft_name}
+%files
 %doc README.md
 %license LICENSE.md
-
-%{pkg_prefix}/%{pkg_libdir}/librocfft.so.0{,.*}
+%{_libdir}/librocfft.so.0{,.*}

 %files devel
-%{pkg_prefix}/include/rocfft/
-%{pkg_prefix}/%{pkg_libdir}/librocfft.so
-%{pkg_prefix}/%{pkg_libdir}/cmake/rocfft/
+%{_includedir}/rocfft/
+%{_libdir}/cmake/rocfft/
+%{_libdir}/librocfft.so

-%if %{with test}
 %files test
-%{pkg_prefix}/bin/rocfft-test
-%{pkg_prefix}/bin/rtc_helper_crash
-%endif
+%{_bindir}/rocfft-test
+%{_bindir}/rtc_helper_crash

 %changelog
-* Mon Jan 26 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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}
@@ -4,13 +4,13 @@ Date: Mon, 3 Nov 2025 06:33:31 -0800
 Subject: [PATCH] rocm-origami remove scope for variables

 ---
- shared/origami/cmake/origami-config.cmake.in | 11 -----------
- 1 file changed, 11 deletions(-)
+ cmake/origami-config.cmake.in | 12 ------------
+ 1 file changed, 12 deletions(-)

-diff --git a/shared/origami/cmake/origami-config.cmake.in b/shared/origami/cmake/origami-config.cmake.in
+diff --git a/cmake/origami-config.cmake.in b/cmake/origami-config.cmake.in
 index d6c8000d0261..19370e7dd5bd 100644
--- a/shared/origami/cmake/origami-config.cmake.in
-+++ b/shared/origami/cmake/origami-config.cmake.in
+--- 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")
@@ -1,118 +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}
-%global upstreamname origami

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-%global origami_name rocm-origami%{pkg_suffix}
+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

-Name:       rocm-origami%{pkg_suffix}
-Version:    %{rocm_version}
-Release:    1%{?dist}
-Summary:    Analytical GEMM Solution Selection
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DCMAKE_VERBOSE_MAKEFILE=ON

-License:    MIT
-URL:        https://github.com/ROCm/rocm-libraries
-Source0:    %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
-# License file is not in the 7.1.0 tag, but is here
-Source2:    https://github.com/ROCm/rocm-libraries/tree/develop/shared/origami/LICENSE.md
-
-#
-# Workaround this hipblaslt build issue
-# CMake Error at /usr/lib64/cmake/origami/origami-config.cmake:11 (message):
+# Workaround hipblaslt build issue:
 #   origami::origami target is missing
-#
-# hipblaslt from rocm-libraries does not use cmake to find origami
 # https://github.com/ROCm/rocm-libraries/issues/2422
-# So they would not have run into this issue.
-Patch1:     0001-rocm-origami-remove-scope-for-variables.patch
+Patch0:         0001-rocm-origami-remove-scope-for-variables.patch

-ExclusiveArch: x86_64 riscv64
-
-BuildRequires: cmake
-BuildRequires: gcc-c++
-BuildRequires: rocm-cmake%{pkg_suffix}
-BuildRequires: rocm-comgr%{pkg_suffix}-devel
-BuildRequires: rocm-llvm%{pkg_suffix}-macros
-BuildRequires: rocm-hip%{pkg_suffix}-devel
-BuildRequires: rocr-runtime%{pkg_suffix}-devel
-
-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
-BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
+BuildRequires:  cmake
+BuildRequires:  cmake(hip)
 BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-BuildRequires:  git
+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.
+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: %{origami_name}%{?_isa} = %{version}-%{release}
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description devel
+%description    devel
 %{summary}

-%prep
-%autosetup -p3 -n %{upstreamname}
-
-# The license file
-cp %{SOURCE2} .
-
+%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

-%build
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix}
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/origami/LICENSE.md

-%cmake_build
-
-%install
-%cmake_install
-
-# Extra license
-rm -f %{buildroot}%{pkg_prefix}/share/doc/origami/LICENSE.md
-
-%files -n %{origami_name}
+%files
 %doc README.md
 %license LICENSE.md
-%{pkg_prefix}/%{pkg_libdir}/liborigami.so.0{,.*}
+%{_libdir}/liborigami.so.0{,.*}

 %files devel
-%{pkg_prefix}/include/origami/
-%{pkg_prefix}/%{pkg_libdir}/cmake/origami/
-%{pkg_prefix}/%{pkg_libdir}/liborigami.so
+%{_includedir}/origami/
+%{_libdir}/cmake/origami/
+%{_libdir}/liborigami.so

 %changelog
-* Mon Feb 23 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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}
@@ -1,131 +1,55 @@
-%global upstreamname rocRAND
-%global rocm_release 7.1
-%global rocm_patch 1
-%global rocm_version %{rocm_release}.%{rocm_patch}
+# 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

-%bcond compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}/
-%global pkg_suffix -%{rocm_release}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%endif
-%global rocrand_name rocrand%{pkg_suffix}
-
-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//' )
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# relevant HW is required to run %check
+# 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
-# enable building of tests if test is enabled
 %if %{with test}
 %global build_test ON
 %else
 %global build_test OFF
 %endif

-# For docs
-%bcond doc 0
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}

-# Use ninja if it is available
-%bcond ninja 1
+# rocm builds with clang
+%global toolchain clang

-%if %{with ninja}
-%global cmake_generator -G Ninja
-%else
-%global cmake_generator %{nil}
-%endif
-
-# The common parts of the cmake configuration
-%global cmake_config \\\
-  -DBUILD_TEST=%build_test \\\
-  -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \\\
-  -DCMAKE_BUILD_TYPE=%build_type \\\
-  -DCMAKE_EXPORT_COMPILE_COMMANDS=%{build_compile_db} \\\
-  -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \\\
-  -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \\\
-  -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \\\
-  -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \\\
-  -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \\\
-  -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \\\
-  -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \\\
-  -DCMAKE_SKIP_RPATH=ON \\\
-  -DROCM_SYMLINK_LIBS=OFF
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
-
-%global gpu_list %{rocm_gpu_list_default}
-%global _gpu_list gfx1100
-
-# export an llvm compilation database
-# Useful for input for other llvm tools
-%bcond export 0
-%if %{with export}
-%global build_compile_db ON
-%else
-%global build_compile_db OFF
-%endif
-
-Name:           rocrand%{pkg_suffix}
+Name:           rocrand
 Version:        %{rocm_version}
-Release:        7%{?dist}
+Release:        %autorelease
 Summary:        ROCm random number generator
-
-Url:            https://github.com/ROCm/rocRAND
 License:        MIT AND BSD-3-Clause
-Source0:        %{url}/archive/rocm-%{version}.tar.gz#/%{upstreamname}-%{version}.tar.gz
+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:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}-modules
-
-%if %{with test} || %{with check}
-BuildRequires:  gtest-devel
+BuildRequires:  cmake(amd_comgr)
+%if %{with test}
+BuildRequires:  cmake(GTest)
 %endif
-
-%if %{with doc}
-BuildRequires:  doxygen
-%endif
-
-%if %{with ninja}
+BuildRequires:  cmake(hip)
+BuildRequires:  cmake(hsa-runtime64)
+BuildRequires:  compiler-rt
+BuildRequires:  lld
+BuildRequires:  llvm
 BuildRequires:  ninja
-%endif
-
-Provides:       rocrand%{pkg_suffix} = %{version}-%{release}
-
-ExclusiveArch:  x86_64 riscv64
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros

 %description
 The rocRAND project provides functions that generate pseudo-random and
@@ -135,75 +59,40 @@ 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
+%package        devel
 Summary:        The rocRAND development package
-Requires:       %{rocrand_name}%{?_isa} = %{version}-%{release}
+Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description devel
+%description    devel
 The rocRAND development package.

 %if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
-Requires:       %{rocrand_name}%{?_isa} = %{version}-%{release}
+Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
 %endif

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/rocrand
-%else
-%autosetup -p1 -n %{upstreamname}-rocm-%{version}
-%endif
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/rocrand/LICENSE.md

-
-# On Tumbleweed Q3,2025
-# https://github.com/ROCm/rocm-libraries/issues/83
-# /usr/include/gtest/internal/gtest-port.h:273:2: error: C++ versions less than C++17 are not supported.
-# Convert the c++11's to c++17
-#sed -i -e 's@set(CMAKE_CXX_STANDARD 11)@set(CMAKE_CXX_STANDARD 17)@' {,test/{cpp_wrapper,package}/}CMakeLists.txt
-
-%build
-
-%cmake %{cmake_generator} %{cmake_config} \
-       -DAMDGPU_TARGETS=%{gpu_list} \
-
-%cmake_build
-
-%install
-
-%cmake_install
-
-rm -f %{buildroot}%{pkg_prefix}/share/doc/rocrand/LICENSE.md
-
-%files -n %{rocrand_name}
+%files
 %doc README.md
 %license LICENSE.md
+%{_libdir}/librocrand.so.1{,.*}

-%if %{with debug}
-%{pkg_prefix}/%{pkg_libdir}/librocrand-d.so.1{,.*}
-%else
-%{pkg_prefix}/%{pkg_libdir}/librocrand.so.1{,.*}
-%endif
-
-%files devel 
-%{pkg_prefix}/include/rocrand/
-%{pkg_prefix}/%{pkg_libdir}/cmake/rocrand/
-%if %{with debug}
-%{pkg_prefix}/%{pkg_libdir}/librocrand-d.so
-%else
-%{pkg_prefix}/%{pkg_libdir}/librocrand.so
-%endif
+%files devel
+%{_includedir}/rocrand/
+%{_libdir}/cmake/rocrand/
+%{_libdir}/librocrand.so

 %if %{with test}
 %files test
-%{pkg_prefix}/bin/test_*
-%{pkg_prefix}/bin/rocRAND/
+%{_bindir}/rocRAND/
+%{_bindir}/test_*
 %endif

 %changelog
-* Mon Jan 26 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%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}
@@ -1,162 +1,108 @@
-%global upstreamname rocthrust
-%global rocm_release 7.1
-%global rocm_patch 1
-%global rocm_version %{rocm_release}.%{rocm_patch}
+# 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 compat 0
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-
-# Compiler is hipcc, which is clang based:
-%global toolchain rocm
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/' -e 's/-mtls-dialect=gnu2//')
-# there is no debug package
-%global debug_package %{nil}
-
-# Option to test suite for testing on real HW:
-%bcond check 0
-%if %{with check}
+# 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

-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-# Threaded compression reduces the build time.
-%global _source_payload w7T0.xzdio
-%global _binary_payload w7T0.xzdio
+%global rocm_release 7.1
+%global rocm_patch 1
+%global rocm_version %{rocm_release}.%{rocm_patch}

-Name:           rocthrust%{pkg_suffix}
+# rocm builds with clang
+%global toolchain clang
+
+Name:           rocthrust
 Version:        %{rocm_version}
-Release:        1%{?dist}
-Summary:        ROCm Thrust libary
+Release:        %autorelease
+Summary:        ROCm Thrust library

-License:        Apache-2.0 AND BSD-2-Clause AND BSD-3-Clause AND BSL-1.0 AND MIT AND LicenseRef-Fedora-Public-Domain
+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 headers files that are Boost 1.0 licensed
+# ./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

-URL:            https://github.com/ROCm/rocm-libraries
-Source0:        %{url}/releases/download/rocm-%{version}/%{upstreamname}.tar.gz#/%{upstreamname}-%{version}.tar.gz
+BuildOption(conf):  -G Ninja
+BuildOption(conf):  -DAMDGPU_TARGETS=%{rocm_gpu_list_default}
+BuildOption(conf):  -DBUILD_TEST=%{build_test}

-BuildRequires:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocprim%{pkg_suffix}-static
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-
-%if %{with check}
-BuildRequires:  gtest-devel
-BuildRequires:  rocminfo
+%if %{with test}
+BuildRequires:  cmake(GTest)
 %endif
-
-ExclusiveArch:  x86_64 riscv64
+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:        The %{upstreamname} development package
-Provides:       %{name}-static = %{version}-%{release}
+%package        devel
+Summary:        Libraries and headers for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description devel
-The %{upstreamname} development package.
+%description    devel
+%{summary}

-%prep
-%if %{with gitcommit}
-%setup -q -n rocm-libraries-%{commit0}
-cd projects/rocthrust
-%else
-%autosetup -n %{upstreamname} -p1
+%if %{with test}
+%package        test
+Summary:        Tests for %{name}
+Requires:       %{name}%{?_isa} = %{version}-%{release}
+
+%description    test
+%{summary}
 %endif

-#
-# The ROCMExportTargetsHeaderOnly.cmake file
-# generates a files that reference the install location of other files
-# Make this change so they match
-sed -i -e 's/ROCM_INSTALL_LIBDIR lib/ROCM_INSTALL_LIBDIR %{pkg_libdir}/' cmake/ROCMExportTargetsHeaderOnly.cmake
+%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

-%build
-%if %{with gitcommit}
-cd projects/rocthrust
-%endif
+%install -a
+rm -f %{buildroot}%{_docdir}/rocthrust/LICENSE

-
-%if %{with check}
-# Building all the gpu's does not make sense
-# Build only the first one, this only works well with rpmbuild.
-gpu=`rocm_agent_enumerator | head -n 1`
-%endif
-
-%cmake \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_LINKER=%rocmllvm_bindir/ld.lld \
-    -DCMAKE_AR=%rocmllvm_bindir/llvm-ar \
-    -DCMAKE_RANLIB=%rocmllvm_bindir/llvm-ranlib \
-    -DBUILD_FILE_REORG_BACKWARD_COMPATIBILITY=OFF \
-    -DBUILD_TEST=%{build_test} \
-%if %{with check}
-    -DAMDGPU_TARGETS=${gpu} \
-%endif
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DROCM_SYMLINK_LIBS=OFF
-
-%cmake_build
-
-%install
-%cmake_install
-
-# Extra license
-rm -f %{buildroot}%{pkg_prefix}/share/doc/rocthrust/LICENSE
-
-%check
-%if %{with check}
-%ctest
-%endif
-
-%files devel
+%files
 %doc README.md
 %license LICENSE
 %license NOTICES.txt
-%{pkg_prefix}/include/thrust
-%{pkg_prefix}/%{pkg_libdir}/cmake/rocthrust/
+
+%files devel
+%{_includedir}/thrust/
+%{_libdir}/cmake/rocthrust/
+
+%if %{with test}
+%files test
+%{_bindir}/test_*
+%{_bindir}/rocthrust/
+%endif

 %changelog
-* Mon Jan 26 2026 Yifan Xu <xuyifan@iscas.ac.cn> - 7.1.1-1
- Import from upstream
+%autochangelog
@@ -1,217 +1,106 @@
-%global upstreamname roctracer
+# 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}

-%bcond_with compat
-%if %{with compat}
-%global pkg_libdir lib
-%global pkg_prefix %{_prefix}/lib64/rocm/rocm-%{rocm_release}
-%global pkg_suffix -%{rocm_release}
-%global pkg_module rocm%{pkg_suffix}
-%else
-%global pkg_libdir %{_lib}
-%global pkg_prefix %{_prefix}
-%global pkg_suffix %{nil}
-%global pkg_module default
-%endif
-
+# rocm stack builds with clang
 %global toolchain clang
-# hipcc does not support some clang flags
-%global build_cxxflags %(echo %{optflags} | sed -e 's/-fstack-protector-strong/-Xarch_host -fstack-protector-strong/' -e 's/-fcf-protection/-Xarch_host -fcf-protection/')

-# Needs ROCm HW and is only suitable for local testing
-# GPU_TARGETS in the cmake config are only for testing
-%bcond test 0
-
-%bcond doc 0
-
-%bcond debug 0
-%if %{with debug}
-%global build_type DEBUG
-%else
-%global build_type RelWithDebInfo
-%endif
-
-# Compression type and level for source/binary package payloads.
-#  "w7T0.xzdio"	xz level 7 using %%{getncpus} threads
-%define _source_payload	w7T0.xzdio
-%define _binary_payload	w7T0.xzdio
-
-Name:           roctracer%{pkg_suffix}
+Name:           roctracer
 Version:        %{rocm_version}
-Release:        1%{?dist}
-Summary:        ROCm Tracer Callback/Activity Library for Performance tracing AMD GPUs
-
-Url:            https://github.com/ROCm/%{upstreamname}
+Release:        %autorelease
+Summary:        ROCm Tracer Callback/Activity Library
+Url:            https://github.com/ROCm/roctracer
+VCS:            git:https://github.com/ROCm/roctracer.git
 License:        MIT
-Source0:        %{url}/archive/rocm-%{rocm_version}.tar.gz#/%{upstreamname}-%{rocm_version}.tar.gz
+#!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:  llvm
-BuildRequires:  llvm-devel
 BuildRequires:  clang
-BuildRequires:  clang-devel
 BuildRequires:  clang-tools-extra
-BuildRequires:  clang-tools-extra-devel
-BuildRequires:  lld
-BuildRequires:  lld-devel
-
-BuildRequires:  hipcc
-BuildRequires:  compiler-rt
-BuildRequires:  rocm-device-libs
-
 BuildRequires:  cmake
-BuildRequires:  gcc-c++
-BuildRequires:  rocm-cmake%{pkg_suffix}
-BuildRequires:  rocm-comgr%{pkg_suffix}-devel
-BuildRequires:  rocm-llvm%{pkg_suffix}-macros
-BuildRequires:  rocm-hip%{pkg_suffix}-devel
-BuildRequires:  rocr-runtime%{pkg_suffix}-devel
-#BuildRequires:  rocm-rpm-macros%{pkg_suffix}
-
-BuildRequires:  libatomic1
-# https://github.com/ROCm/roctracer/issues/113
+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)
-
-%if %{with doc}
-BuildRequires:  doxygen
-BuildRequires:  texlive-adjustbox
-BuildRequires:  texlive-dvips
-BuildRequires:  texlive-ec
-BuildRequires:  texlive-hanging
-BuildRequires:  texlive-latex
-BuildRequires:  texlive-makeindex
-BuildRequires:  texlive-metafont
-BuildRequires:  texlive-multirow
-BuildRequires:  texlive-newunicodechar
-BuildRequires:  texlive-stackengine
-BuildRequires:  texlive-texlive-scripts
-BuildRequires:  texlive-tocloft
-BuildRequires:  texlive-ulem
-BuildRequires:  texlive-url
-BuildRequires:  texlive-wasy
-BuildRequires:  texlive-wasysym
-%endif
-
-ExclusiveArch:  x86_64 riscv64
+BuildRequires:  rocm-cmake
+BuildRequires:  rocm-device-libs
+BuildRequires:  rocm-llvm-macros

 %description
-ROC-tracer
+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.

-* ROC-tracer library: Runtimes Generic Callback/Activity APIs
-
-  The goal of the implementation is to provide a generic independent
-  from specific runtime profiler to trace API and asynchronous activity.
-
-  The API provides functionality for registering the runtimes API
-  callbacks and asynchronous activity records pool support.
-
-* ROC-TX library: Code Annotation Events API
-
-  Includes API for:
-
-  * roctxMark
-  * roctxRangePush
-  * roctxRangePop
-
-%post -p /sbin/ldconfig
-%postun -p /sbin/ldconfig
-
-%package devel
-Summary:        The %{name} development package
+%package        devel
+Summary:        The roctracer development package
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description devel
-The headers of libraries for %{name}.
-
-%if %{with doc}
-%package doc
-Summary:        Docs for %{name}
-
-%description doc
-%{summary}
-%endif
+%description    devel
+The roctracer development package.

 %if %{with test}
-%package test
+%package        test
 Summary:        Tests for %{name}
 Requires:       %{name}%{?_isa} = %{version}-%{release}

-%description test
+%description    test
 %{summary}
 %endif

-%prep
-%autosetup -p1 -n %{upstreamname}-rocm-%{version}
-
+%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/@../%{pkg_libdir}/@' test/run.sh
-
+sed -i -e 's@../lib/@../%{_lib}/@' test/run.sh
 %endif

-%build
-%cmake \
-    -DCMAKE_BUILD_TYPE=%{build_type} \
-    -DCMAKE_C_COMPILER=%rocmllvm_bindir/clang \
-    -DCMAKE_CXX_COMPILER=%rocmllvm_bindir/clang++ \
-    -DCMAKE_CXX_FLAGS="-I%{pkg_prefix}/include"\
-    -DCMAKE_EXE_LINKER_FLAGS="-L %{pkg_prefix}/%{pkg_libdir} -lamdhip64" \
-    -DCMAKE_INSTALL_LIBDIR=%{pkg_libdir} \
-    -DCMAKE_INSTALL_PREFIX=%{pkg_prefix} \
-    -DCMAKE_MODULE_PATH=%{pkg_prefix}/%{pkg_libdir}/cmake/hip \
-    -DCMAKE_PREFIX_PATH=%{rocmllvm_cmakedir}/.. \
-    -DROCM_SYMLINK_LIBS=OFF \
-    -DGPU_TARGETS=%{rocm_gpu_list_test} \
-    -DHIP_PLATFORM=amd \
-    -DHIP_HIPCC_FLAGS="-I%{pkg_prefix}/include -L%{pkg_prefix}/%{pkg_libdir} -lamdhip64" \
-    -DBUILD_SHARED_LIBS=ON
-
-%cmake_build
-
-%if %{with doc}
-%cmake_build -t doc
-%endif
-
-%install
-%cmake_install
-
-# Only install the pdf
-rm -rf rm %{buildroot}%{pkg_prefix}/share/html
-# Extra licenses
-# Fedora
-rm -f %{buildroot}%{pkg_prefix}/share/doc/*/LICENSE.md
-# OpenSUSE
-rm -f %{buildroot}%{pkg_prefix}/share/doc/*/*/LICENSE.md
-
+%install -a
+rm -f %{buildroot}%{_datadir}/doc/%{name}/LICENSE.md
+rm -rf %{buildroot}%{_datadir}/doc/%{name}-asan

 %files
 %license LICENSE.md
 %doc README.md
-%{pkg_prefix}/%{pkg_libdir}/libroctracer64.so.*
-%{pkg_prefix}/%{pkg_libdir}/libroctx64.so.*
-%{pkg_prefix}/%{pkg_libdir}/roctracer/
+%{_libdir}/libroctracer64.so.*
+%{_libdir}/libroctx64.so.*
+%{_libdir}/roctracer/

 %files devel
-%{pkg_prefix}/include/roctracer
-%{pkg_prefix}/%{pkg_libdir}/libroctracer64.so
-%{pkg_prefix}/%{pkg_libdir}/libroctx64.so
-
-%if %{with doc}
-%files doc
-%{pkg_prefix}/share/doc/roctracer/
-%endif
+%{_includedir}/roctracer/
+%{_libdir}/libroctracer64.so
+%{_libdir}/libroctx64.so

 %if %{with test}
 %files test
-%{pkg_prefix}/share/roctracer/
+%{_datadir}/roctracer/
 %endif

 %changelog
-* Mon Jan 26 2026 Yifan Xu <xuyifan@iscas.ac.cn> -  - 7.1.1-1
- Import from upstream
+%autochangelog
				`@@ -0,0 +1 @@`
				`u ollama - "Runs Ollama" /var/lib/ollama /sbin/nologin`