rocmPackages: 7.2.3 -> 10.1.0

nixos/doc/manual/release-notes/rl-2611.section.md

@@ -188,6 +188,25 @@
   authentication layer before exposing it to other users.
 - `services.autobrr.secretFile` has been removed, as autobrr no longer uses a session secret since version 1.82.0. Remove the option from your configuration.
 
+- `rocmPackages` has been updated from ROCm 7.2.4 to ROCm 10.1.0. AMD skipped
+  the 8.x and 9.x series. ROCm 10 breaks ABI and API:
+  `amdsmi` moves to the 27.0.0 ABI (`libamd_smi.so.27`, several clock/throttle
+  APIs removed), `rocsparse` drops `rocsparse_indextype_u16`, `hipcub` removes
+  its deprecated iterators and operators, and `miopen` drops its OpenCL backend.
+  Update out-of-tree packages that compile against `rocmPackages` for these
+  changes. The ROCm LLVM toolchain is now LLVM 24.
+
+  `rocmPackages.rocmlir` now builds from `ROCm/rocmlirTriton`, the rocMLIR
+  fork that AMDMIGraphX 10.1 pins, instead of `ROCm/rocMLIR`.
+
+  `rocmPackages.rocm-bandwidth-test` has been removed; upstream retired it.
+
+  `composable_kernel`, the `miopen` CK solvers, and `vllm` no longer support
+  `gfx900`/`gfx906` (Vega / MI25 / MI50 / Radeon VII) on ROCm 10.
+  `rocblas`/`hipblas`-only consumers such as `llama-cpp` and `ollama` still
+  build for those targets. To stay on ROCm 7.2.x, use the NixOS 26.05
+  stable channel.
+
 - Artalk has been updated to 2.10.0. Its default configuration and data
   directory discovery changed; see the [upstream migration
   guide](https://artalk.js.org/en/guide/releases/v2.10.0.html) when invoking

pkgs/by-name/bl/blender/package.nix

@@ -165,8 +165,12 @@ stdenv'.mkDerivation (finalAttrs: {
                 '${lib.getLib brotli}/lib/libbrotlidec.dylib'
   '')
   + (lib.optionalString rocmSupport ''
-    substituteInPlace extern/hipew/src/hipew.c --replace-fail '"/opt/rocm/hip/lib/libamdhip64.so.${lib.versions.major rocmPackages.clr.version}"' '"${rocmPackages.clr}/lib/libamdhip64.so"'
-    substituteInPlace extern/hipew/src/hipew.c --replace-fail '"opt/rocm/hip/bin"' '"${rocmPackages.clr}/bin"'
+    # Point hipew's bare "libamdhip64.so" candidate at clr. The HIP soname
+    # major (libamdhip64.so.7) no longer matches the ROCm release major, so
+    # keying the path off clr.version breaks on ROCm 10.
+    substituteInPlace extern/hipew/src/hipew.c \
+      --replace-fail '"libamdhip64.so"' '"${rocmPackages.clr}/lib/libamdhip64.so"' \
+      --replace-fail '"opt/rocm/hip/bin"' '"${rocmPackages.clr}/bin"'
   '')
   + (lib.optionalString oneapiSupport ''
     substituteInPlace intern/cycles/kernel/device/oneapi/CMakeLists.txt \

pkgs/by-name/ol/ollama/package.nix

@@ -261,11 +261,27 @@ goBuild (finalAttrs: {
       # (cuda_v12 / cuda_v13 / rocm_v7_1 / rocm_v7_2 / vulkan / cuda_jetpack*).
       rocmMajorVersion = lib.versions.major rocmPackages.clr.version;
       rocmMinorVersion = lib.versions.minor rocmPackages.clr.version;
+      # cmake/local.cmake only defines ROCm backend labels rocm_v7_1 (Windows)
+      # and rocm_v7_2 (Linux); the two "currently share build settings" and the
+      # label is an upstream *interface tag*, deliberately decoupled from the
+      # actual ROCm toolchain version (which comes from rocmPackages via
+      # buildInputs + AMDGPU_TARGETS). Our rocmPackages is now ROCm 10, so the
+      # derived rocm_v10_0 would hit `Unknown ROCm backend` (no such preset).
+      # Map any ROCm >= 8 to the Linux rocm_v7_2 build path — the GGML_HIP
+      # runner it produces is version-agnostic and links our ROCm 10 libs.
+      # TODO: file an upstream ollama PR adding an explicit rocm_v10_0
+      # preset (llama/server/CMakePresets.json) + local.cmake branch, then drop
+      # this remap.
+      rocmLinuxBackend =
+        if lib.versionAtLeast rocmMajorVersion "8" then
+          "rocm_v7_2"
+        else
+          "rocm_v${rocmMajorVersion}_${rocmMinorVersion}";
       llamaBackend =
         if enableCuda then
           "cuda_v${cudaMajorVersion}"
         else if enableRocm then
-          "rocm_v${rocmMajorVersion}_${rocmMinorVersion}"
+          rocmLinuxBackend
         else if enableVulkan then
           "vulkan"
         else

pkgs/by-name/op/opensplat/package.nix

@@ -3,6 +3,8 @@
   stdenv,
   cmake,
   ninja,
+  pkg-config,
+  libdrm,
   fetchFromGitHub,
   python3,
   opencv,
@@ -81,6 +83,10 @@ stdenv'.mkDerivation (finalAttrs: {
   ]
   ++ lib.optionals rocmSupport [
     rocmPackages.clr
+    # ROCm 10's rocm_smi-config.cmake does find_dependency(PkgConfig) (rocm-smi-lib
+    # now requires libdrm_amdgpu.pc), pulled in transitively via find_package(Torch)
+    # -> Caffe2 LoadHIP -> rocm_smi, so pkg-config must be on the host path.
+    pkg-config
   ]
   ++ lib.optionals cudaSupport [
     cudaPackages.cuda_nvcc
@@ -106,8 +112,10 @@ stdenv'.mkDerivation (finalAttrs: {
     rocmPackages.hipblaslt
     rocmPackages.rocblas
     rocmPackages.rocsolver
-    rocmPackages.rocsolver
     rocmPackages.miopen
+    # rocm_smi-config.cmake (pulled via find_package(Torch)->Caffe2 LoadHIP) runs
+    # pkg_check_modules(libdrm) in ROCm 10, so libdrm.pc must be on PKG_CONFIG_PATH.
+    libdrm
   ]
   ++ lib.optionals cudaSupport [
     cudaPackages.cuda_cudart

pkgs/development/compilers/llvm/common/openmp/default.nix

@@ -24,98 +24,140 @@
 
 assert lib.assertMsg (ompdGdbSupport -> ompdSupport) "OMPD GDB support requires OMPD support!";
 
-stdenv.mkDerivation (finalAttrs: {
-  pname = "openmp";
-  inherit version;
-
-  src =
-    if monorepoSrc != null then
-      runCommand "openmp-src-${version}"
-        {
-          inherit (monorepoSrc) passthru;
-          strictDeps = true;
-          __structuredAttrs = true;
-        }
-        ''
-          mkdir -p "$out"
-          cp -r ${monorepoSrc}/cmake "$out"
-          cp -r ${monorepoSrc}/openmp "$out"
-        ''
-    else
-      src;
-
-  sourceRoot = "${finalAttrs.src.name}/openmp";
-
-  outputs = [
-    "out"
-    "dev"
-  ];
-
-  patches =
-    lib.optional (lib.versionOlder release_version "19") (getVersionFile "openmp/fix-find-tool.patch")
-    ++ [
-      (getVersionFile "openmp/run-lit-directly.patch")
+let
+  # LLVM 23 removed the legacy standalone OpenMP build (cmake <llvm>/openmp now
+  # aborts with a FATAL_ERROR). OpenMP must be built through the runtimes entry
+  # point (cmake <llvm>/runtimes -DLLVM_ENABLE_RUNTIMES=openmp), which supplies
+  # the project() call and the LLVM cmake modules. Older LLVMs keep the
+  # standalone layout.
+  useRuntimesBuild = lib.versionAtLeast release_version "23";
+in
+stdenv.mkDerivation (
+  finalAttrs:
+  {
+    pname = "openmp";
+    inherit version;
+
+    src =
+      if monorepoSrc != null then
+        runCommand "openmp-src-${version}"
+          {
+            inherit (monorepoSrc) passthru;
+            strictDeps = true;
+            __structuredAttrs = true;
+          }
+          (
+            if useRuntimesBuild then
+              ''
+                mkdir -p "$out/llvm"
+                cp -r ${monorepoSrc}/cmake "$out"
+                cp -r ${monorepoSrc}/runtimes "$out"
+                cp -r ${monorepoSrc}/openmp "$out"
+                cp -r ${monorepoSrc}/llvm/cmake "$out/llvm/"
+                # openmp/runtime/unittests unconditionally pulls in gtest from
+                # ../third-party/unittest, so it must be present even with tests off.
+                cp -r ${monorepoSrc}/third-party "$out"
+              ''
+            else
+              ''
+                mkdir -p "$out"
+                cp -r ${monorepoSrc}/cmake "$out"
+                cp -r ${monorepoSrc}/openmp "$out"
+              ''
+          )
+      else
+        src;
+
+    sourceRoot = "${finalAttrs.src.name}/${if useRuntimesBuild then "runtimes" else "openmp"}";
+
+    outputs = [
+      "out"
+      "dev"
     ];
 
-  nativeBuildInputs = [
-    cmake
-    python3
-    perl
-    ninja
-    pkg-config
-    lit
-  ];
-
-  propagatedNativeBuildInputs = [
-    checkPhaseThreadLimitHook
-  ];
-
-  buildInputs = [
-    llvm
-  ]
-  ++ lib.optionals (ompdSupport && ompdGdbSupport) [
-    python3
-  ];
-
-  strictDeps = true;
-
-  cmakeFlags = [
-    (lib.cmakeBool "LIBOMP_ENABLE_SHARED" (
-      !stdenv.hostPlatform.isStatic && stdenv.hostPlatform.hasSharedLibraries
-    ))
-    (lib.cmakeBool "LIBOMP_OMPD_SUPPORT" ompdSupport)
-    (lib.cmakeBool "LIBOMP_OMPD_GDB_SUPPORT" ompdGdbSupport)
-    (lib.cmakeFeature "CLANG_TOOL" "${clang-unwrapped}/bin/clang")
-    (lib.cmakeFeature "OPT_TOOL" "${llvm}/bin/opt")
-    (lib.cmakeFeature "LINK_TOOL" "${llvm}/bin/llvm-link")
-  ]
-  ++ devExtraCmakeFlags;
-
-  doCheck = false;
-
-  checkTarget = "check-openmp";
-
-  preCheck = ''
-    patchShebangs ../tools/archer/tests/deflake.bash
-  '';
-
-  __structuredAttrs = true;
-
-  meta = llvm_meta // {
-    homepage = "https://openmp.llvm.org/";
-    description = "Support for the OpenMP language";
-    longDescription = ''
-      The OpenMP subproject of LLVM contains the components required to build an
-      executable OpenMP program that are outside the compiler itself.
-      Contains the code for the runtime library against which code compiled by
-      "clang -fopenmp" must be linked before it can run and the library that
-      supports offload to target devices.
-    '';
-    # "All of the code is dual licensed under the MIT license and the UIUC
-    # License (a BSD-like license)":
-    license = with lib.licenses; [
-      mit
-      ncsa
+    patches =
+      lib.optional (lib.versionOlder release_version "19") (getVersionFile "openmp/fix-find-tool.patch")
+      ++ [
+        (getVersionFile "openmp/run-lit-directly.patch")
+      ];
+
+    nativeBuildInputs = [
+      cmake
+      python3
+      perl
+      ninja
+      pkg-config
+      lit
+    ];
+
+    propagatedNativeBuildInputs = [
+      checkPhaseThreadLimitHook
+    ];
+
+    buildInputs = [
+      llvm
+    ]
+    ++ lib.optionals (ompdSupport && ompdGdbSupport) [
+      python3
     ];
-  };
-})
+
+    strictDeps = true;
+
+    cmakeFlags = [
+      (lib.cmakeBool "LIBOMP_ENABLE_SHARED" (
+        !stdenv.hostPlatform.isStatic && stdenv.hostPlatform.hasSharedLibraries
+      ))
+      (lib.cmakeBool "LIBOMP_OMPD_SUPPORT" ompdSupport)
+      (lib.cmakeBool "LIBOMP_OMPD_GDB_SUPPORT" ompdGdbSupport)
+      (lib.cmakeFeature "CLANG_TOOL" "${clang-unwrapped}/bin/clang")
+      (lib.cmakeFeature "OPT_TOOL" "${llvm}/bin/opt")
+      (lib.cmakeFeature "LINK_TOOL" "${llvm}/bin/llvm-link")
+    ]
+    ++ lib.optionals useRuntimesBuild [
+      # Build through the runtimes entry point (see useRuntimesBuild above).
+      (lib.cmakeFeature "LLVM_ENABLE_RUNTIMES" "openmp")
+      # We disable tests (doCheck = false); avoid pulling in third-party/unittest,
+      # which the trimmed source tree does not contain.
+      (lib.cmakeBool "LLVM_INCLUDE_TESTS" false)
+    ]
+    ++ devExtraCmakeFlags;
+
+    doCheck = false;
+
+    checkTarget = "check-openmp";
+
+    preCheck = ''
+      patchShebangs ../tools/archer/tests/deflake.bash
+    '';
+
+    __structuredAttrs = true;
+
+    meta = llvm_meta // {
+      homepage = "https://openmp.llvm.org/";
+      description = "Support for the OpenMP language";
+      longDescription = ''
+        The OpenMP subproject of LLVM contains the components required to build an
+        executable OpenMP program that are outside the compiler itself.
+        Contains the code for the runtime library against which code compiled by
+        "clang -fopenmp" must be linked before it can run and the library that
+        supports offload to target devices.
+      '';
+      # "All of the code is dual licensed under the MIT license and the UIUC
+      # License (a BSD-like license)":
+      license = with lib.licenses; [
+        mit
+        ncsa
+      ];
+    };
+  }
+  # Kept conditional so that derivations for older LLVMs stay unchanged.
+  // lib.optionalAttrs useRuntimesBuild {
+    # The runtimes build writes generated headers (omp.h, ...) back into the
+    # openmp/ source tree, which is a sibling of sourceRoot and therefore left
+    # read-only by unpackPhase (only sourceRoot gets `chmod u+w`). Make the whole
+    # checkout writable.
+    preConfigure = ''
+      chmod -R u+w ..
+    '';
+  }
+)

pkgs/development/python-modules/torch/source/default.nix

@@ -55,6 +55,7 @@
   openssl,
   numactl,
   llvmPackages,
+  libdrm,
 
   # dependencies
   filelock,
@@ -399,6 +400,16 @@ buildPythonPackage.override { inherit stdenv; } (finalAttrs: {
       --replace-fail "list(APPEND ATen_HIP_INCLUDE \''${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/include)" "" \
       --replace-fail "list(APPEND ATen_HIP_INCLUDE \''${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/library/include)" ""
   ''
+  + lib.optionalString rocmSupport ''
+    # ROCm 10's rocprofiler-sdk ships an OMPT plugin whose one-time init dlopens a
+    # library while libomp holds its (non-recursive) init ticket lock: OpenBLAS's
+    # constructor queries the CPU count -> libomp __kmp_middle_initialize ->
+    # ompt_pre_init -> rocprofiler-sdk -> dlopen -> OpenBLAS ctor -> libomp again,
+    # busy-waiting on the held lock forever, so `import torch` hangs (surfaced as a
+    # pythonImportsCheck hang; also hits runtime, e.g. vllm). Disable the OMPT tool
+    # interface by default (setdefault, so a user wanting rocprofv3 can override).
+    sed -i '/^    if USE_GLOBAL_DEPS:$/i\    os.environ.setdefault("OMP_TOOL", "disabled")' torch/__init__.py
+  ''
   # Remove PyTorch's FindCUDAToolkit.cmake and use CMake's default.
   # NOTE: Parts of pytorch rely on unmaintained FindCUDA.cmake with custom patches to support e.g.
   # newer architectures (sm_90a). We do want to delete vendored patches, but have to keep them
@@ -625,6 +636,13 @@ buildPythonPackage.override { inherit stdenv; } (finalAttrs: {
   ++ lib.optionals rocmSupport [
     rocmtoolkit_joined
     rocmPackages.clr # Added separately so setup hook applies
+    # ROCm 10's rocm-smi exports a rocm_smi-config.cmake that does
+    # pkg_check_modules(_ROCM_SMI_DRM REQUIRED libdrm) (its public headers pull in
+    # libdrm/drm.h via kfd_ioctl.h). rocm-smi propagates libdrm, but torch consumes
+    # it through the rocmtoolkit_joined symlinkJoin, which drops propagatedBuildInputs
+    # -> libdrm.pc is off PKG_CONFIG_PATH and torch's LoadHIP find_package(rocm_smi)
+    # fails at configure. Add libdrm directly so the pkg-config hook finds libdrm.pc.
+    libdrm
   ];
 
   dependencies = [

pkgs/development/python-modules/torchcodec/default.nix

@@ -15,6 +15,7 @@
   libjpeg,
   libpng,
   libwebp,
+  libdrm,
 
   # build-system
   cmake,
@@ -113,7 +114,12 @@ buildPythonPackage.override { inherit (torch) stdenv; } (finalAttrs: {
       libnpp # nppicc
       libnvjpeg # nvjpeg.h
     ]
-  );
+  )
+  ++ lib.optionals rocmSupport [
+    # find_package(Torch) -> Caffe2 LoadHIP -> find_package(rocm_smi), whose ROCm 10
+    # rocm_smi-config.cmake runs pkg_check_modules(libdrm), so libdrm.pc must be present.
+    libdrm
+  ];
 
   build-system = [
     cmake