[CI] Add CK (independent of miopen) full build and test to TheRock.

.github/workflows/build_portable_linux_pytorch_wheels.yml

@@ -175,9 +175,8 @@ jobs:
           ./external-builds/pytorch/pytorch_vision_repo.py checkout --require-related-commit
           ./external-builds/pytorch/pytorch_triton_repo.py checkout
 
-      - name: Create pip cache directory
-        run: mkdir -p /tmp/pipcache
-
+      # Note: determine_version.py sets optional_build_prod_arguments in
+      # GITHUB_ENV, which includes --rocm-sdk-version and --version-suffix.
       - name: Determine optional arguments passed to `build_prod_wheels.py`
         if: ${{ inputs.rocm_version }}
         run: |
@@ -192,7 +191,6 @@ jobs:
           ./external-builds/pytorch/build_prod_wheels.py \
             build \
             --install-rocm \
-            --pip-cache-dir /tmp/pipcache \
             --index-url "${{ inputs.cloudfront_url }}/${{ inputs.amdgpu_family }}/" \
             --clean \
             --output-dir ${{ env.PACKAGE_DIST_DIR }} ${{ env.optional_build_prod_arguments }}
@@ -310,7 +308,7 @@ jobs:
             --include "torch-${TORCH_VERSION}-${CP_VERSION}-linux_x86_64.whl" \
             --include "torchaudio-${TORCHAUDIO_VERSION}-${CP_VERSION}-linux_x86_64.whl" \
             --include "torchvision-${TORCHVISION_VERSION}-${CP_VERSION}-linux_x86_64.whl" \
-            --include "triton-${TRITON_VERSION}-${CP_VERSION}-linux_x86_64.whl"
+            --include "triton-${TRITON_VERSION}-${CP_VERSION}-linux_x86_64.whl" \
             --include "apex-${APEX_VERSION}-${CP_VERSION}-linux_x86_64.whl"
 
       - name: (Re-)Generate Python package release index

.github/workflows/build_portable_linux_pytorch_wheels_ci.yml

@@ -0,0 +1,133 @@
+# CI variant of build_portable_linux_pytorch_wheels.yml.
+#
+# Key differences from the release workflow:
+#   - Installs ROCm packages via --find-links (CI artifacts) instead of
+#     --index-url (release index)
+#   - Builds torch only (no torchvision, torchaudio, or triton)
+#   - No S3 upload, staging, testing, or promotion — just build + sanity check
+#   - (Not yet implemented) this can use an explict separate sccache bucket
+#
+# TODO(#3291): Build more packages (torchvision, torchaudio, triton, etc.)
+# TODO(#3291): Upload packages to S3 (via upload_python_packages.py?)
+#
+# Both workflows share build_prod_wheels.py for the actual build logic.
+# See https://github.com/ROCm/TheRock/issues/3291 for convergence plans.
+
+name: Build Portable Linux PyTorch Wheels (CI)
+
+on:
+  workflow_call:
+    inputs:
+      artifact_group:
+        type: string
+        required: true
+      python_version:
+        type: string
+        default: "3.12"
+      pytorch_git_ref:
+        description: PyTorch ref to checkout (typically "release/X.Y")
+        type: string
+        default: "release/2.10"
+      rocm_package_find_links_url:
+        description: URL for pip --find-links to install ROCm packages
+        type: string
+        required: true
+      rocm_version:
+        description: ROCm package version to install and build against (e.g. 7.10.0.dev0)
+        type: string
+        required: true
+  workflow_dispatch:
+    inputs:
+      artifact_group:
+        description: "The artifact group to build (e.g. gfx94X-dcgpu, gfx120X-all)"
+        type: string
+        default: gfx94X-dcgpu
+      python_version:
+        type: string
+        default: "3.12"
+      pytorch_git_ref:
+        description: PyTorch ref to checkout (typically "release/X.Y")
+        type: string
+        default: "release/2.10"
+      rocm_package_find_links_url:
+        description: URL for pip --find-links to install ROCm packages
+        type: string
+        required: true
+      rocm_version:
+        description: ROCm package version to install and build against (e.g. 7.10.0.dev0)
+        type: string
+        required: true
+
+permissions:
+  contents: read
+
+run-name: Build portable Linux PyTorch Wheels CI (${{ inputs.artifact_group }}, py${{ inputs.python_version }}, ${{ inputs.pytorch_git_ref }})
+
+jobs:
+  build_pytorch_wheels:
+    name: Build PyTorch | ${{ inputs.artifact_group }} | torch ${{ inputs.pytorch_git_ref }} | py${{ inputs.python_version }}
+    runs-on: ${{ github.repository_owner == 'ROCm' && 'azure-linux-scale-rocm' || 'ubuntu-24.04' }}
+    container:
+      image: ghcr.io/rocm/therock_build_manylinux_x86_64@sha256:db2b63f938941dde2abc80b734e64b45b9995a282896d513a0f3525d4591d6cb
+    env:
+      PACKAGE_DIST_DIR: ${{ github.workspace }}/output/packages/dist
+      optional_build_prod_arguments: ""
+    steps:
+      - name: Checkout
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Configure Git Identity
+        run: |
+          git config --global user.name "therockbot"
+          git config --global user.email "therockbot@amd.com"
+
+      - name: Select Python version
+        run: |
+          python build_tools/github_actions/python_to_cp_version.py \
+            --python-version ${{ inputs.python_version }}
+
+      - name: Add selected Python version to PATH
+        run: |
+          python_dir="/opt/python/${{ env.cp_version }}"
+          if ! [ -x "${python_dir}/bin/python" ]; then
+            echo "ERROR: Could not find python: ${python_dir}"
+            exit 1
+          fi
+          echo "${python_dir}/bin" >> "$GITHUB_PATH"
+
+      - name: Checkout PyTorch source (nightly)
+        if: ${{ inputs.pytorch_git_ref == 'nightly' }}
+        run: |
+          ./external-builds/pytorch/pytorch_torch_repo.py checkout \
+            --repo-hashtag nightly
+
+      - name: Checkout PyTorch source (stable)
+        if: ${{ inputs.pytorch_git_ref != 'nightly' }}
+        run: |
+          ./external-builds/pytorch/pytorch_torch_repo.py checkout \
+            --gitrepo-origin https://github.com/ROCm/pytorch.git \
+            --repo-hashtag ${{ inputs.pytorch_git_ref }}
+
+      # Note: determine_version.py sets optional_build_prod_arguments in
+      # GITHUB_ENV, which includes --rocm-sdk-version and --version-suffix.
+      - name: Determine optional arguments passed to `build_prod_wheels.py`
+        if: ${{ inputs.rocm_version }}
+        run: |
+          pip install packaging
+          python build_tools/github_actions/determine_version.py \
+            --rocm-version ${{ inputs.rocm_version }}
+
+      - name: Build PyTorch wheels
+        run: |
+          ./external-builds/pytorch/build_prod_wheels.py \
+            build \
+            --install-rocm \
+            --find-links "${{ inputs.rocm_package_find_links_url }}" \
+            --clean \
+            --output-dir ${{ env.PACKAGE_DIST_DIR }} \
+            ${{ env.optional_build_prod_arguments }}
+
+      - name: Sanity check wheel
+        run: |
+          python external-builds/pytorch/sanity_check_wheel.py \
+            ${{ env.PACKAGE_DIST_DIR }}/

.github/workflows/build_windows_pytorch_wheels.yml

@@ -195,6 +195,8 @@ jobs:
               --torch-dir ${{ env.CHECKOUT_ROOT }}/torch \
               --require-related-commit
 
+      # Note: determine_version.py sets optional_build_prod_arguments in
+      # GITHUB_ENV, which includes --rocm-sdk-version and --version-suffix.
       - name: Determine optional arguments passed to `build_prod_wheels.py`
         if: ${{ inputs.rocm_version }}
         run: |

.github/workflows/ci.yml

@@ -102,6 +102,7 @@ jobs:
       rocm_package_version: ${{ needs.setup.outputs.rocm_package_version }}
       test_type: ${{ needs.setup.outputs.test_type }}
       sanity_check_only_for_family: ${{ matrix.variant.sanity_check_only_for_family == true }}
+      build_pytorch: ${{ matrix.variant.build_pytorch == true }}
     permissions:
       contents: read
       id-token: write

.github/workflows/ci_linux.yml

@@ -32,6 +32,9 @@ on:
         type: string
       sanity_check_only_for_family:
         type: boolean
+      build_pytorch:
+        type: boolean
+        default: false
 
 permissions:
   contents: read
@@ -164,3 +167,25 @@ jobs:
       package_find_links_url: ${{ needs.build_portable_linux_python_packages.outputs.package_find_links_url }}
       python_version: "3.12"
       rocm_version: ${{ inputs.rocm_package_version }}
+
+  # TODO(#3291): Add test_pytorch_wheels job
+  build_portable_linux_pytorch_wheels_ci:
+    needs: [build_portable_linux_python_packages]
+    name: Build PyTorch
+    if: >-
+      ${{
+        !failure() &&
+        !cancelled() &&
+        (
+          inputs.use_prebuilt_artifacts == 'false' ||
+          inputs.use_prebuilt_artifacts == 'true'
+        ) &&
+        inputs.build_pytorch == true
+      }}
+    uses: ./.github/workflows/build_portable_linux_pytorch_wheels_ci.yml
+    with:
+      artifact_group: ${{ inputs.artifact_group }}
+      python_version: "3.12"
+      pytorch_git_ref: "release/2.10"
+      rocm_package_find_links_url: ${{ needs.build_portable_linux_python_packages.outputs.package_find_links_url }}
+      rocm_version: ${{ inputs.rocm_package_version }}

.github/workflows/ci_nightly.yml

@@ -85,6 +85,7 @@ jobs:
       rocm_package_version: ${{ needs.setup.outputs.rocm_package_version }}
       test_type: ${{ needs.setup.outputs.test_type }}
       sanity_check_only_for_family: ${{ matrix.variant.sanity_check_only_for_family == true }}
+      build_pytorch: ${{ matrix.variant.build_pytorch == true }}
     permissions:
       contents: read
       id-token: write

.github/workflows/ci_windows.yml

@@ -164,3 +164,6 @@ jobs:
       package_find_links_url: ${{ needs.build_windows_python_packages.outputs.package_find_links_url }}
       python_version: "3.12"
       rocm_version: ${{ inputs.rocm_package_version }}
+
+  # TODO(#3291): Add build_windows_pytorch_wheels_ci
+  # TODO(#3291): Add test_pytorch_wheels job

BUILD_TOPOLOGY.toml

@@ -137,6 +137,7 @@ artifact_groups = ["third-party-sysdeps", "base"]
 description = "Compiler, runtimes, and core profiling"
 artifact_groups = [
     "compiler",
+    "core-amdsmi",
     "core-runtime",
     "third-party-libs",
     "hip-runtime",
@@ -208,6 +209,12 @@ type = "generic"
 artifact_group_deps = ["base", "third-party-sysdeps"]
 source_sets = ["rocm-systems"]
 
+[artifact_groups.core-amdsmi]
+description = "AMD SMI tool (amdsmi)"
+type = "generic"
+artifact_group_deps = ["base", "third-party-sysdeps"]
+source_sets = ["rocm-systems"]
+
 [artifact_groups.compiler]
 description = "AMD LLVM toolchain and compiler infrastructure"
 type = "generic"
@@ -267,9 +274,9 @@ source_sets = ["rocm-systems"]  # rocprofiler-sdk is in rocm-systems
 [artifact_groups.dctools-core]
 description = "Data center management tools with minimal dependencies"
 type = "generic"
-artifact_group_deps = ["core-runtime", "profiler-core"]
+artifact_group_deps = ["core-runtime", "core-amdsmi", "profiler-core"]
 # TODO: rocm-systems included for projects/hip/VERSION (see CMakeLists.txt)
-source_sets = ["base", "rocm-systems"]  # RDC uses amdsmi from base
+source_sets = ["base", "rocm-systems"]  # RDC uses amdsmi from core-amdsmi
 
 # Future artifact groups
 # [artifact_groups.dctools-rocm]
@@ -280,7 +287,7 @@ source_sets = ["base", "rocm-systems"]  # RDC uses amdsmi from base
 [artifact_groups.profiler-apps]
 description = "Profiler applications and analysis tools"
 type = "generic"
-artifact_group_deps = ["profiler-core", "compiler"]
+artifact_group_deps = ["profiler-core", "compiler", "core-amdsmi"]
 source_sets = ["rocm-systems", "profiler-extras"]  # rocprofiler-systems + trace decoder
 
 [artifact_groups.iree-libs]
@@ -437,6 +444,14 @@ feature_name = "CORE_RUNTIME"
 feature_group = "CORE"
 disable_platforms = ["windows"]
 
+[artifacts.core-amdsmi]
+artifact_group = "core-amdsmi"
+type = "target-neutral"
+artifact_deps = ["base", "sysdeps"]
+feature_name = "CORE_AMDSMI"
+feature_group = "CORE"
+disable_platforms = ["windows"]
+
 [artifacts.core-hip]
 artifact_group = "hip-runtime"
 type = "target-neutral"
@@ -454,7 +469,7 @@ feature_group = "CORE"  # Part of core, enabled by default
 [artifacts.rocrtst]
 artifact_group = "rocrtst"
 type = "target-neutral"
-artifact_deps = ["core-runtime", "core-ocl"]
+artifact_deps = ["core-runtime", "core-ocl", "sysdeps-hwloc"]
 feature_name = "CORE_RUNTIME_TESTS"
 feature_group = "CORE"
 disable_platforms = ["windows"]
@@ -479,7 +494,7 @@ feature_group = "CORE"  # Part of core, enabled by default
 [artifacts.blas]
 artifact_group = "math-libs"
 type = "target-specific"
-artifact_deps = ["core-runtime", "core-hip", "host-blas", "host-suite-sparse", "rocprofiler-sdk"]
+artifact_deps = ["core-runtime", "core-hip", "core-amdsmi", "host-blas", "host-suite-sparse", "rocprofiler-sdk"]
 split_databases = ["rocblas", "hipblaslt"]
 
 [artifacts.fft]
@@ -528,7 +543,7 @@ artifact_deps = []
 [artifacts.composable-kernel]
 artifact_group = "ml-libs"
 type = "target-specific"
-artifact_deps = ["core-runtime", "core-hip"]
+artifact_deps = ["core-runtime", "core-hip", "rand", "rocprofiler-sdk"]
 
 [artifacts.miopen]
 artifact_group = "ml-libs"
@@ -568,7 +583,7 @@ disable_platforms = ["windows"]
 [artifacts.rccl]
 artifact_group = "comm-libs"
 type = "target-specific"
-artifact_deps = ["core-runtime", "core-hip", "hipify", "rocprofiler-sdk"]
+artifact_deps = ["core-runtime", "core-hip", "hipify", "rocprofiler-sdk", "core-amdsmi"]
 disable_platforms = ["windows"]
 
 # --- Profiler Tools ---
@@ -598,7 +613,7 @@ disable_platforms = ["windows"]
 [artifacts.rocprofiler-systems]
 artifact_group = "profiler-apps"
 type = "target-neutral"
-artifact_deps = ["amd-llvm", "core-hip", "rocprofiler-sdk"]
+artifact_deps = ["amd-llvm", "core-hip", "rocprofiler-sdk", "core-amdsmi"]
 feature_name = "ROCPROFSYS"
 feature_group = "PROFILER"
 disable_platforms = ["windows"]
@@ -608,7 +623,7 @@ disable_platforms = ["windows"]
 [artifacts.rdc]
 artifact_group = "dctools-core"
 type = "target-neutral"
-artifact_deps = ["core-hip", "rocprofiler-sdk", "sysdeps"]
+artifact_deps = ["core-hip", "rocprofiler-sdk", "sysdeps", "core-amdsmi"]
 feature_group = "DC_TOOLS"  # Part of DC tools, enabled by default via THEROCK_ENABLE_ALL
 disable_platforms = ["windows"]
 

base/CMakeLists.txt

@@ -56,37 +56,6 @@ therock_cmake_subproject_provide_package(rocm-core rocm-core lib/cmake/rocm-core
 therock_cmake_subproject_activate(rocm-core)
 
 
-################################################################################
-# amdsmi
-################################################################################
-
-if(NOT WIN32)  # TODO(#36): Enable on Windows and/or make subproject inclusion generally optional
-
-therock_cmake_subproject_declare(amdsmi
-  EXTERNAL_SOURCE_DIR "${THEROCK_ROCM_SYSTEMS_SOURCE_DIR}/projects/amdsmi"
-  BINARY_DIR "${CMAKE_CURRENT_BINARY_DIR}/amdsmi"
-  USE_DIST_AMDGPU_TARGETS
-  BACKGROUND_BUILD
-  CMAKE_ARGS
-    "-DCMAKE_VERBOSE_MAKEFILE=OFF"
-    -DBUILD_TESTS=${THEROCK_BUILD_TESTING}
-  BUILD_DEPS
-    therock-googletest
-  RUNTIME_DEPS
-    rocm-core
-    ${THEROCK_BUNDLED_LIBDRM}
-  INSTALL_RPATH_DIRS
-    lib
-  INTERFACE_INSTALL_RPATH_DIRS
-    lib
-)
-therock_cmake_subproject_provide_package(amdsmi
-  amd_smi lib/cmake)
-therock_cmake_subproject_activate(amdsmi)
-list(APPEND _optional_artifact_deps amdsmi)
-
-endif()
-
 
 ################################################################################
 # rocm_smi_lib