Move CUDA CI coverage to 13.0

.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)

.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() {
@@ -76,14 +77,19 @@ 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

.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 ==="
 

.github/workflows/cuda-perf.yml

@@ -110,7 +110,7 @@ jobs:
       secrets-env: EXECUTORCH_HF_TOKEN
       runner: ${{ contains(matrix.model, 'Qwen3.5-35B-A3B') && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
-      gpu-arch-version: "12.6"
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       upload-artifact: model-${{ matrix.model_safe }}-${{ matrix.quant }}
@@ -158,7 +158,7 @@ jobs:
       timeout: 90
       runner: ${{ contains(matrix.model, 'Qwen3.5-35B-A3B') && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
-      gpu-arch-version: "12.6"
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       download-artifact: model-${{ matrix.model_safe }}-${{ matrix.quant }}

.github/workflows/cuda-windows.yml

@@ -64,7 +64,7 @@ jobs:
       secrets-env: EXECUTORCH_HF_TOKEN
       runner: linux.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
-      gpu-arch-version: 12.8
+      gpu-arch-version: "13.0"
       docker-image: ci-image:executorch-ubuntu-22.04-cuda-windows
       submodules: recursive
       upload-artifact: ${{ matrix.model_repo }}-${{ matrix.model_name }}-cuda-windows-${{ matrix.quant }}
@@ -146,7 +146,7 @@ jobs:
       timeout: 240
       runner: windows.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
-      gpu-arch-version: 12.8
+      gpu-arch-version: "13.0"
       download-artifact: ${{ matrix.model_repo }}-${{ matrix.model_name }}-cuda-windows-${{ matrix.quant }}
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
       script: |
@@ -158,7 +158,7 @@ jobs:
           \$ErrorActionPreference = 'Stop'
           \$PSNativeCommandUseErrorActionPreference = \$true
 
-          \$env:CUDA_HOME = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.8'
+          \$env:CUDA_HOME = 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.0'
           \$env:CUDA_PATH = \$env:CUDA_HOME
           \$env:PATH = \"\$env:CUDA_HOME\bin;\$env:PATH\"
           nvcc --version
@@ -169,5 +169,5 @@ jobs:
             throw 'RUNNER_ARTIFACT_DIR is empty. Ensure download-artifact is configured for windows_job.yml.'
           }
 
-          .ci/scripts/test_model_e2e_windows.ps1 -Device cuda-windows -HfModel '${{ matrix.model_repo }}/${{ matrix.model_name }}' -QuantName '${{ matrix.quant }}' -ModelDir \$artifactDir -ExpectedCudaVersion '12.8'
+          .ci/scripts/test_model_e2e_windows.ps1 -Device cuda-windows -HfModel '${{ matrix.model_repo }}/${{ matrix.model_name }}' -QuantName '${{ matrix.quant }}' -ModelDir \$artifactDir -ExpectedCudaVersion '13.0'
         }"

.github/workflows/cuda.yml

@@ -1,6 +1,6 @@
 # Test ExecuTorch CUDA Build Compatibility
 # This workflow tests whether ExecuTorch can be successfully built with CUDA support
-# across different CUDA versions (12.6, 12.8, 12.9, 13.0) using the command:
+# across different CUDA versions (12.6, 13.0) using the command:
 #   ./install_executorch.sh
 #
 # Note: ExecuTorch automatically detects the system CUDA version using nvcc and
@@ -31,7 +31,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        cuda-version: ["12.6", "12.8", "12.9", "13.0"]
+        cuda-version: ["12.6", "13.0"]
 
     name: test-executorch-cuda-build-${{ matrix.cuda-version }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
@@ -66,7 +66,7 @@ jobs:
             echo "CUDA build results: ${{ needs.test-cuda-builds.result }}"
             exit 1
           else
-            echo "SUCCESS: All ExecuTorch CUDA builds (12.6, 12.8, 12.9, 13.0) completed successfully!"
+            echo "SUCCESS: All ExecuTorch CUDA builds (12.6, 13.0) completed successfully!"
           fi
 
   test-models-cuda:
@@ -79,7 +79,7 @@ jobs:
       timeout: 90
       runner: linux.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
-      gpu-arch-version: 12.6
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -114,7 +114,7 @@ jobs:
       timeout: 90
       runner: linux.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
-      gpu-arch-version: 12.6
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
@@ -271,7 +271,7 @@ jobs:
       secrets-env: EXECUTORCH_HF_TOKEN
       runner: ${{ (matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' || matrix.model.name == 'gemma-4-31B-it-HQQ-INT4') && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
-      gpu-arch-version: 12.6
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       upload-artifact: ${{ matrix.model.repo }}-${{ matrix.model.name }}-cuda-${{ matrix.quant }}
@@ -406,7 +406,7 @@ jobs:
       timeout: 90
       runner: ${{ (matrix.model.name == 'Qwen3.5-35B-A3B-HQQ-INT4' || matrix.model.name == 'gemma-4-31B-it-HQQ-INT4') && 'linux.aws.a100' || 'linux.g5.4xlarge.nvidia.gpu' }}
       gpu-arch-type: cuda
-      gpu-arch-version: 12.6
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       download-artifact: ${{ matrix.model.repo }}-${{ matrix.model.name }}-cuda-${{ matrix.quant }}
@@ -442,7 +442,7 @@ jobs:
       download-artifact: ${{ matrix.artifact }}
       runner: linux.g5.4xlarge.nvidia.gpu
       gpu-arch-type: cuda
-      gpu-arch-version: 12.6
+      gpu-arch-version: "13.0"
       use-custom-docker-registry: false
       submodules: recursive
       ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}