Update (base update)

[ghstack-poisoned]

Author: SS-JIA

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-release/2.12
+release/2.13

.ci/docker/common/install_pytorch.sh

@@ -16,6 +16,21 @@ install_domains() {
   pip_install --no-build-isolation --user "git+https://github.com/pytorch/vision.git@${TORCHVISION_VERSION}"
 }
 
+configure_pytorch_compiler() {
+  local cxx_version
+  cxx_version=$(/usr/bin/c++ --version | head -n1)
+  if [[ "${cxx_version}" == *clang* ]]; then
+    local clang_major
+    clang_major=$(/usr/bin/c++ -dumpversion | cut -d. -f1)
+    if [[ "${clang_major}" =~ ^[0-9]+$ && "${clang_major}" -lt 16 ]] &&
+      command -v gcc >/dev/null &&
+      command -v g++ >/dev/null; then
+      export CC=gcc
+      export CXX=g++
+    fi
+  fi
+}
+
 install_pytorch_and_domains() {
   git clone https://github.com/pytorch/pytorch.git
 
@@ -33,14 +48,15 @@ install_pytorch_and_domains() {
   if [[ -n "${PYTORCH_BUILD_MAX_JOBS:-}" ]]; then
     export MAX_JOBS="${PYTORCH_BUILD_MAX_JOBS}"
   fi
+  configure_pytorch_compiler
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"
 
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.27
+  TORCHVISION_VERSION=release/0.28
   export TORCHVISION_VERSION
 
   install_domains

.ci/scripts/setup-windows.ps1

@@ -14,6 +14,10 @@ conda activate et
 # Install test dependencies
 pip install -r .ci/docker/requirements-ci.txt
 
+# PyTorch 2.13 headers use designated initializers in c10, which require C++20
+# with MSVC.
+$env:CMAKE_ARGS = "$env:CMAKE_ARGS -DCMAKE_CXX_STANDARD=20"
+
 # The Windows CI image ships CUDA toolkits on PATH, so install_executorch
 # (setup.py) auto-enables EXECUTORCH_BUILD_CUDA whenever the detected nvcc
 # version is in SUPPORTED_CUDA_VERSIONS. CPU-only jobs install CPU torch, so a

.ci/scripts/test_model_e2e_windows.ps1

@@ -172,22 +172,34 @@ try {
             "-DCUDAToolkit_ROOT=$env:CUDA_HOME"
         )
     }
+    $cmakeCommonArgs = @("-DCMAKE_CXX_STANDARD=20")
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build ExecuTorch (CUDA)"
     $numCores = [Math]::Max([Environment]::ProcessorCount - 1, 1)
+    $cmakeOut = Join-Path -Path $executorchRoot -ChildPath "cmake-out"
+    $executorchCmakeDir = Join-Path -Path $cmakeOut -ChildPath "lib\cmake\ExecuTorch"
+    $executorchConfig = Join-Path -Path $executorchCmakeDir -ChildPath "executorch-config.cmake"
     # EXECUTORCH_BUILD_EXTENSION_IMAGE defaults OFF in the preset, so it must be
     # enabled explicitly here (matching the Makefile CUDA target used on Linux).
     # The dinov2 runner links the installed extension_image.lib; without this the
     # main install never builds it and dinov2_runner fails to link (LNK1181).
-    cmake --preset llm-release-cuda -DEXECUTORCH_BUILD_EXTENSION_IMAGE=ON @cmakeCudaArgs
-    cmake --build cmake-out --target install --config Release -j $numCores
+    cmake --preset llm-release-cuda -DEXECUTORCH_BUILD_EXTENSION_IMAGE=ON @cmakeCommonArgs @cmakeCudaArgs
+    cmake --build $cmakeOut --target install --config Release -j $numCores
+    if (-not (Test-Path -Path $executorchConfig -PathType Leaf)) {
+        throw "ExecuTorch CMake package config not found after install: $executorchConfig"
+    }
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build $runnerTarget"
+    $cmakePackageArgs = @(
+        "-DCMAKE_FIND_ROOT_PATH=$cmakeOut",
+        "-DCMAKE_PREFIX_PATH=$cmakeOut;$executorchCmakeDir",
+        "-Dexecutorch_DIR=$executorchCmakeDir"
+    )
     Push-Location (Join-Path -Path $executorchRoot -ChildPath "examples\models\$runnerPath")
     try {
-        cmake --preset $runnerPreset @cmakeCudaArgs
+        cmake --preset $runnerPreset @cmakeCommonArgs @cmakePackageArgs @cmakeCudaArgs
         cmake --build (Join-Path -Path $executorchRoot -ChildPath "cmake-out\examples\models\$runnerPath") --target $runnerTarget --config Release -j $numCores
     }
     finally {

.ci/scripts/utils.sh

@@ -107,7 +107,7 @@ install_pytorch_and_domains() {
   local torch_release=$(cat version.txt)
   # Download key must match the upload key below (basename of dist/*.whl,
   # which always carries setup.py's resolved +gitHASH). Branch-ref pins
-  # like `release/2.12` would otherwise produce `+gitrelease` here and
+  # like `release/2.13` would otherwise produce `+gitrelease` here and
   # never hit the cache.
   local torch_short_hash=$(git rev-parse --short=7 HEAD)
   local torch_wheel_path="cached_artifacts/pytorch/executorch/pytorch_wheels/${system_name}/${python_version}"
@@ -178,7 +178,7 @@ install_pytorch_and_domains() {
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.27
+  TORCHVISION_VERSION=release/0.28
   export TORCHVISION_VERSION
 
   install_domains

.github/workflows/_unittest.yml

@@ -24,6 +24,10 @@ on:
         required: false
         type: string
         default: '3.10'
+      linux-timeout:
+        required: false
+        type: number
+        default: 90
 
 jobs:
   linux:
@@ -36,7 +40,7 @@ jobs:
       docker-image: ${{ inputs.docker-image }}
       submodules: 'recursive'
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-      timeout: 90
+      timeout: ${{ inputs.linux-timeout }}
       script: |
         set -eux
         .ci/scripts/unittest-linux.sh --build-tool "${{ inputs.build-tool }}" --build-mode "${{ inputs.build-mode }}" --editable "${{ inputs.editable }}"

.github/workflows/cuda.yml

@@ -211,7 +211,7 @@ jobs:
 
         # Run Gemma 4 31B tests (quant unit tests + pipeline integration tests)
         pip install gguf
-        python -m pytest examples/models/gemma4_31b/quant/tests/ examples/models/gemma4_31b/tests/ --ignore=examples/models/gemma4_31b/tests/test_mlx_pipeline.py -v -o "addopts="
+        python -m pytest examples/models/gemma4_31b/tests/ --ignore=examples/models/gemma4_31b/tests/test_mlx_pipeline.py -v -o "addopts="
 
   unittest-cuda-runtime:
     name: unittest-cuda-runtime

.github/workflows/mlx.yml

@@ -85,7 +85,6 @@ jobs:
           backends/mlx/test/test_serialization_dedup.py \
           backends/mlx/test/test_slot_recycling.py \
           backends/mlx/test/test_sample.py \
-          examples/models/gemma4_31b/quant/tests/test_pack_mlx.py \
           examples/models/gemma4_31b/tests/test_mlx_pipeline.py \
           -v
         echo "::endgroup::"
@@ -158,10 +157,10 @@ jobs:
           --prompt-len 4 \
           --max-new-tokens 5 2>&1)
         echo "$OUTPUT"
-        if echo "$OUTPUT" | grep -q "Generated token ids: \[167, 94, 253, 88, 227\]"; then
+        if echo "$OUTPUT" | grep -q "Generated token ids: \[167, 180, 167, 180, 167\]"; then
           echo "Success: Qwen 3.5 MoE MLX export + inference completed with expected output"
         else
-          echo "Failed: unexpected output (expected [167, 94, 253, 88, 227])"
+          echo "Failed: unexpected output (expected [167, 180, 167, 180, 167])"
           exit 1
         fi
         echo "::endgroup::"

.github/workflows/trunk.yml

@@ -1006,6 +1006,7 @@ jobs:
       build-mode: Release
       build-tool: cmake
       docker-image: ci-image:executorch-ubuntu-22.04-clang12
+      linux-timeout: 120
 
   llm-server:
     needs: [changed-files, run-decision]

CMakeLists.txt

@@ -79,6 +79,10 @@ project(executorch
         VERSION "${ET_VERSION_MAJOR}.${ET_VERSION_MINOR}.${ET_VERSION_PATCH}"
 )
 
+if(WIN32)
+  add_compile_definitions(NOMINMAX)
+endif()
+
 # Hexagon toolchain with release build complains about code in third party
 # libraries. Must come after project(), which runs the toolchain file that sets
 # CMAKE_SYSTEM_PROCESSOR, and before add_subdirectory(third-party).