Reapply PyTorch 2.12 pin with CI fixes (pytorch#19699)

Summary
Trying again! This time working more on the cuda 

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils
@Sebastian-Larsson @robell @rascani

Author: JacobSzwejbka

.ci/docker/build.sh

@@ -81,7 +81,7 @@ case "${IMAGE_NAME}" in
     LINTRUNNER=""
     GCC_VERSION=11
     CUDA_WINDOWS_CROSS_COMPILE=yes
-    CUDA_VERSION=12.8
+    CUDA_VERSION=13.0
     SKIP_PYTORCH=yes
     ;;
   executorch-ubuntu-24.04-gcc14)
@@ -97,6 +97,10 @@ esac
 TORCH_VERSION=$(cat ci_commit_pins/pytorch.txt)
 BUILD_DOCS=1
 
+if [[ "${GCC_VERSION:-}" == "11" && -z "${SKIP_PYTORCH:-}" ]]; then
+  PYTORCH_BUILD_MAX_JOBS=6
+fi
+
 # Copy requirements-lintrunner.txt from root to here
 cp ../../requirements-lintrunner.txt ./
 
@@ -109,6 +113,7 @@ docker build \
   --build-arg "PYTHON_VERSION=${PYTHON_VERSION}" \
   --build-arg "MINICONDA_VERSION=${MINICONDA_VERSION}" \
   --build-arg "TORCH_VERSION=${TORCH_VERSION}" \
+  --build-arg "PYTORCH_BUILD_MAX_JOBS=${PYTORCH_BUILD_MAX_JOBS:-}" \
   --build-arg "BUCK2_VERSION=${BUCK2_VERSION}" \
   --build-arg "LINTRUNNER=${LINTRUNNER:-}" \
   --build-arg "BUILD_DOCS=${BUILD_DOCS}" \

.ci/docker/ci_commit_pins/pytorch.txt

@@ -1 +1 @@
-release/2.11
+release/2.12

.ci/docker/common/install_cache.sh

@@ -76,6 +76,9 @@ init_sccache() {
   # This is the remote cache bucket
   export SCCACHE_BUCKET=ossci-compiler-cache-circleci-v2
   export SCCACHE_S3_KEY_PREFIX=executorch
+  export SCCACHE_REGION=us-east-1
+  export AWS_REGION=us-east-1
+  export AWS_DEFAULT_REGION=us-east-1
   export SCCACHE_IDLE_TIMEOUT=0
   export SCCACHE_ERROR_LOG=/tmp/sccache_error.log
   export RUST_LOG=sccache::server=error

.ci/docker/common/install_cuda.sh

@@ -10,10 +10,10 @@
 
 set -ex
 
-# CUDA version must be specified (e.g., 12.8)
+# CUDA version must be specified (e.g., 13.0)
 CUDA_VERSION="${CUDA_VERSION:?CUDA_VERSION must be set}"
 
-# Convert version format (e.g., 12.8 -> 12-8 for package names)
+# Convert version format (e.g., 13.0 -> 13-0 for package names)
 CUDA_VERSION_DASH=$(echo "${CUDA_VERSION}" | tr '.' '-')
 
 # Add NVIDIA package repository

.ci/docker/common/install_cuda_windows_cross_compile.sh

@@ -17,6 +17,7 @@ declare -A CUDA_DRIVER_MAP=(
     ["12.6"]="12.6.3:561.17"
     ["12.8"]="12.8.1:572.61"
     ["12.9"]="12.9.1:576.57"
+    ["13.0"]="13.0.2:"
 )
 
 install_mingw() {
@@ -76,19 +77,26 @@ install_windows_cuda() {
     CUDA_VERSION=$(echo "${CUDA_INFO}" | cut -d: -f1)
     CUDA_DRIVER_VERSION=$(echo "${CUDA_INFO}" | cut -d: -f2)
 
-    echo "Using CUDA ${CUDA_VERSION} with driver ${CUDA_DRIVER_VERSION}"
+    if [ -n "${CUDA_DRIVER_VERSION}" ]; then
+        echo "Using CUDA ${CUDA_VERSION} with driver ${CUDA_DRIVER_VERSION}"
+        CUDA_INSTALLER="cuda_${CUDA_VERSION}_${CUDA_DRIVER_VERSION}_windows.exe"
+    else
+        echo "Using CUDA ${CUDA_VERSION}"
+        CUDA_INSTALLER="cuda_${CUDA_VERSION}_windows.exe"
+    fi
 
     echo "Installing Windows CUDA toolkit ${CUDA_VERSION}..."
 
     mkdir -p "${INSTALL_DIR}"
     cd "${INSTALL_DIR}"
 
-    CUDA_INSTALLER="cuda_${CUDA_VERSION}_${CUDA_DRIVER_VERSION}_windows.exe"
     CUDA_URL="https://developer.download.nvidia.com/compute/cuda/${CUDA_VERSION}/local_installers/${CUDA_INSTALLER}"
 
     # Check if already downloaded and extracted
     if [ -d "${INSTALL_DIR}/extracted/cuda_cudart" ]; then
         echo "Windows CUDA toolkit already installed, skipping download..."
+        chmod -R a+rX "${INSTALL_DIR}"
+        chmod -R a+rwX "${INSTALL_DIR}/extracted/cuda_cudart/cudart/lib"
         return 0
     fi
 
@@ -98,8 +106,11 @@ install_windows_cuda() {
     echo "Extracting CUDA toolkit..."
     7z x "${CUDA_INSTALLER}" -o"extracted" -y
 
-    # Fix permissions so ci-user can access the files
+    # Fix permissions so ci-user can access the files. PyTorch Inductor also
+    # needs to write a MinGW import library beside cudart.lib during Windows
+    # cross-compilation.
     chmod -R a+rX "${INSTALL_DIR}"
+    chmod -R a+rwX "${INSTALL_DIR}/extracted/cuda_cudart/cudart/lib"
 
     # Clean up installer to save space
     rm -f "${CUDA_INSTALLER}"

.ci/docker/common/install_pytorch.sh

@@ -27,14 +27,20 @@ install_pytorch_and_domains() {
   chown -R ci-user .
 
   export _GLIBCXX_USE_CXX11_ABI=1
+  if [[ "$(uname -m)" == "aarch64" ]]; then
+    export BUILD_IGNORE_SVE_UNAVAILABLE=1
+  fi
+  if [[ -n "${PYTORCH_BUILD_MAX_JOBS:-}" ]]; then
+    export MAX_JOBS="${PYTORCH_BUILD_MAX_JOBS}"
+  fi
   # Then build and install PyTorch
   conda_run python setup.py bdist_wheel
   pip_install "$(echo dist/*.whl)"
 
   # Grab the pinned audio and vision commits from PyTorch
   TORCHAUDIO_VERSION=release/2.11
   export TORCHAUDIO_VERSION
-  TORCHVISION_VERSION=release/0.26
+  TORCHVISION_VERSION=release/0.27
   export TORCHVISION_VERSION
 
   install_domains

.ci/docker/ubuntu/Dockerfile

@@ -62,9 +62,12 @@ RUN bash ./install_cache.sh && rm install_cache.sh utils.sh
 ENV SCCACHE_BUCKET ossci-compiler-cache-circleci-v2
 ENV SCCACHE_S3_KEY_PREFIX executorch
 ENV SCCACHE_REGION us-east-1
+ENV AWS_REGION us-east-1
+ENV AWS_DEFAULT_REGION us-east-1
 
 ARG TORCH_VERSION
 ARG SKIP_PYTORCH
+ARG PYTORCH_BUILD_MAX_JOBS
 COPY ./common/install_pytorch.sh install_pytorch.sh
 COPY ./common/utils.sh utils.sh
 RUN if [ -z "${SKIP_PYTORCH}" ]; then bash ./install_pytorch.sh; fi && rm install_pytorch.sh utils.sh

.ci/scripts/export_model_artifact.sh

@@ -518,7 +518,7 @@ fi
 
 DEVICE_ARG=""
 if [ "$DEVICE" = "cuda" ] || [ "$DEVICE" = "cuda-windows" ]; then
-  DEVICE_ARG="--device cuda"
+  DEVICE_ARG="--device cuda:0"
 elif [ "$DEVICE" = "metal" ]; then
   DEVICE_ARG="--device mps"
 fi

.ci/scripts/test-cuda-build.sh

@@ -7,7 +7,7 @@
 
 set -exu
 
-CUDA_VERSION=${1:-"12.6"}
+CUDA_VERSION=${1:-"13.0"}
 
 echo "=== Testing ExecuTorch CUDA ${CUDA_VERSION} Build ==="
 

.ci/scripts/test_model_e2e_windows.ps1

@@ -159,18 +159,31 @@ try {
         }
         Write-Host "CUDA version check passed: $actualCudaVersion"
     }
+    $cmakeCudaArgs = @()
+    if (-not [string]::IsNullOrWhiteSpace($env:CUDA_HOME)) {
+        $cudaNvcc = Join-Path -Path $env:CUDA_HOME -ChildPath "bin\nvcc.exe"
+        if (-not (Test-Path -Path $cudaNvcc -PathType Leaf)) {
+            throw "CUDA compiler not found at '$cudaNvcc'"
+        }
+        $env:CUDACXX = $cudaNvcc
+        $cmakeCudaArgs = @(
+            "-T", "cuda=$env:CUDA_HOME",
+            "-DCMAKE_CUDA_COMPILER=$cudaNvcc",
+            "-DCUDAToolkit_ROOT=$env:CUDA_HOME"
+        )
+    }
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build ExecuTorch (CUDA)"
     $numCores = [Math]::Max([Environment]::ProcessorCount - 1, 1)
-    cmake --preset llm-release-cuda
+    cmake --preset llm-release-cuda @cmakeCudaArgs
     cmake --build cmake-out --target install --config Release -j $numCores
     Write-Host "::endgroup::"
 
     Write-Host "::group::Build $runnerTarget"
     Push-Location (Join-Path -Path $executorchRoot -ChildPath "examples\models\$runnerPath")
     try {
-        cmake --preset $runnerPreset
+        cmake --preset $runnerPreset @cmakeCudaArgs
         cmake --build (Join-Path -Path $executorchRoot -ChildPath "cmake-out\examples\models\$runnerPath") --target $runnerTarget --config Release -j $numCores
     }
     finally {