[ROCm] Test ROCm failures on PR#1554

.github/workflows/linux_rocm.yaml

@@ -72,7 +72,7 @@ jobs:
       fail-fast: false
       matrix:
         python-version: ['3.10']
-        rocm-version: ['7.1']
+        rocm-version: ['7.2']
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     permissions:
       id-token: write
@@ -144,6 +144,32 @@ jobs:
         bash packaging/install_rocjpeg.sh
         echo '::endgroup::'
 
+        echo '::group::VA-API and DRM diagnostics'
+        # libva libraries present in the container
+        ldconfig -p | grep -E "libva|libdrm" || true
+        # DRM render nodes - permissions and group ownership
+        ls -la /dev/dri/ || true
+        # Current user and groups (container user must be in video/render group)
+        id
+        # Check whether mesa-amdgpu-va-drivers was actually installed
+        # (install_rocjpeg.sh silently falls back and skips it if AMD repo is absent)
+        rpm -q mesa-amdgpu-va-drivers libva-amdgpu rocjpeg 2>&1 || true
+        # Find VA driver .so files - if amdgpu_drv_video.so is missing, vaInit will fail
+        find /usr -name "*drv_video*" 2>/dev/null || true
+        # Try vainfo on every render node (no-sudo: true so we can't install vainfo)
+        if command -v vainfo &>/dev/null; then
+          for node in /dev/dri/renderD*; do
+            echo "--- vainfo on ${node} ---"
+            vainfo --display drm --device "${node}" 2>&1 || true
+          done
+        else
+          echo "vainfo not installed, skipping VA-API smoke test"
+        fi
+        # Show which VA driver is selected
+        echo "LIBVA_DRIVER_NAME=${LIBVA_DRIVER_NAME:-<not set>}"
+        echo "LIBVA_DRIVERS_PATH=${LIBVA_DRIVERS_PATH:-<not set>}"
+        echo '::endgroup::'
+
         echo '::group::Install torchcodec from the wheel'
         python -m pip install "${RUNNER_ARTIFACT_DIR}"/*.whl -vvv
         echo '::endgroup::'

packaging/install_rocjpeg.sh

@@ -20,6 +20,11 @@ set -euo pipefail
 # test runners). On the plain build image that repo is absent; there we only need
 # to *compile* against rocjpeg.h/librocjpeg.so, so fall back to installing just
 # those with --nodeps (base libva provides the libva.so.2 soname we link).
+#
+# ROCm >= 7.14 distributes the full ROCm stack (including rocJPEG) as pip wheels
+# (_rocm_sdk_core / _rocm_sdk_devel site-packages). In that case librocjpeg.so
+# and rocjpeg.h are already present and the dnf packages don't exist, so we
+# skip the install entirely.
 install_rocjpeg_build_only() {
     dnf install -y --refresh libva
     dnf install -y "dnf-command(download)" >/dev/null 2>&1 || dnf install -y dnf-plugins-core
@@ -28,5 +33,16 @@ install_rocjpeg_build_only() {
     rpm -Uvh --nodeps "${rpm_dir}"/rocjpeg*.rpm
 }
 
-dnf install -y --refresh rocjpeg-devel libva-amdgpu mesa-amdgpu-va-drivers \
-    || install_rocjpeg_build_only
+# Check if librocjpeg is already available (e.g. via ROCm 7.14+ pip wheels).
+if python3 -c "
+import glob, sys
+# _rocm_sdk_core and _rocm_sdk_devel are the pip-wheel-based ROCm installs
+hits = (glob.glob('/opt/conda/**/librocjpeg.so*', recursive=True) +
+        glob.glob('/opt/rocm/lib/librocjpeg.so*'))
+sys.exit(0 if hits else 1)
+" 2>/dev/null; then
+    echo "librocjpeg already present (ROCm pip-wheel install); skipping dnf install."
+else
+    dnf install -y --refresh rocjpeg-devel libva-amdgpu mesa-amdgpu-va-drivers \
+        || install_rocjpeg_build_only
+fi