ci: bug fixes

Author: ggorman

.github/workflows/pytest-gpu.yml

@@ -70,7 +70,7 @@ jobs:
             --init --network=host
             --device=/dev/kfd --device=/dev/dri
             --ipc=host
-            --group-add video --group-add $(getent group render | cut -d: -f3)
+            --group-add video --group-add "$(getent group render | cut -d: -f3)"
             --cap-add=SYS_PTRACE --security-opt seccomp=unconfined
             --rm -t
             --name ${CONTAINER_BASENAME}
@@ -81,7 +81,6 @@ jobs:
 
       - name: Set per-runner tags
         run: |
-          echo "RUNNER_NAME=$RUNNER_NAME" >> $GITHUB_ENV
           echo "DOCKER_IMAGE=${{ matrix.name }}-$RUNNER_NAME" >> $GITHUB_ENV
           echo "CONTAINER_BASENAME=testrun-${{ matrix.name }}-$RUNNER_NAME" >> $GITHUB_ENV
 
@@ -101,40 +100,43 @@ jobs:
 
       - name: Probe gpu
         run: |
-          # Run a simple driver cmd first (nvidia-smi / rocm-smi)
+          # Make sure CUDA_VISIBLE_DEVICES is at least *something* on NVIDIA
+          # runners; fall back to "all" so the driver probe does not fail.
+          if [[ "${{ matrix.runner_label }}" == "nvidiagpu" && -z "${CUDA_VISIBLE_DEVICES:-}" ]]; then
+            echo "CUDA_VISIBLE_DEVICES=all" >> $GITHUB_ENV
+          fi
+
+          # Run a simple driver-probe command (nvidia-smi / rocm-smi)
           docker run ${{ matrix.flags }} "${DOCKER_IMAGE}" ${{ matrix.test_drive_cmd }}
 
       - name: Test with pytest
         env:
           # Exported earlier in the job; needed inside the container for codecov
           CODECOV_TOKEN: ${{ env.CODECOV_TOKEN }}
         run: |
-          # ------------------------------------------------------------------
           # 1.  Add Codecov’s environment variables (GITHUB_SHA, etc.)
-          # ------------------------------------------------------------------
           ci_env=$(bash <(curl -s https://codecov.io/env))
 
-          # ------------------------------------------------------------------
-          # 2.  Make sure the runner really gave us a GPU mask
-          # ------------------------------------------------------------------
-          : "${CUDA_VISIBLE_DEVICES:?CUDA_VISIBLE_DEVICES not set}"
-
-          # ------------------------------------------------------------------
-          # 3.  Run the test suite inside the image we just built
-          # ------------------------------------------------------------------
-          docker run \
-            --init --rm -t \
-            --name "${CONTAINER_BASENAME}" \
-            --env CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES}" \
-            --gpus "device=${CUDA_VISIBLE_DEVICES}" \
+          # 2.  For NVIDIA, make sure the runner gave us a GPU mask
+          # Default to empty
+          NVIDIA_ENV_FLAGS=""
+          # For NVIDIA, check for device mask AND set specific env vars
+          if [[ "${{ matrix.runner_label }}" == "nvidiagpu" ]]; then
+            : "${CUDA_VISIBLE_DEVICES:?CUDA_VISIBLE_DEVICES not set on NVIDIA runner}"
+            NVIDIA_ENV_FLAGS=" \
+              -e CUDA_LAUNCH_BLOCKING=1 \
+              -e NV_ACC_NOTIFY=3 \
+              -e NV_ACC_DEBUG=1"
+          fi
+
+          # 3.  Run the test suite using the matrix-defined flags
+          docker run ${{ matrix.flags }} \
             ${ci_env} \
             -e CI=true \
             -e PYTHONFAULTHANDLER=1 \
-            -e CUDA_LAUNCH_BLOCKING=1 \
-            -e NV_ACC_NOTIFY=3 \
-            -e NV_ACC_DEBUG=1 \
             -e DEVITO_LOGGING=DEBUG \
             -e CODECOV_TOKEN \
+            ${NVIDIA_ENV_FLAGS} \
             "${DOCKER_IMAGE}" \
             pytest -vv -ra -l -s --full-trace --maxfail=1 \
                    --cov --cov-config=.coveragerc --cov-report=xml \