diff --git a/.github/actionlint.yaml b/.github/actionlint.yaml
new file mode 100644
index 00000000..981dc725
--- /dev/null
+++ b/.github/actionlint.yaml
@@ -0,0 +1,12 @@
+self-hosted-runner:
+  labels:
+    - linux-aarch64-a2-1
+    - linux-aarch64-a2-2
+    - linux-aarch64-a2-4
+    - linux-aarch64-a2-8
+    - linux-aarch64-a2-16
+    - linux-aarch64-a3-1
+    - linux-aarch64-a3-2
+    - linux-aarch64-a3-4
+    - linux-aarch64-a3-8
+    - linux-aarch64-a3-16
diff --git a/.github/workflows/probe-self-hosted-runners.yml b/.github/workflows/probe-self-hosted-runners.yml
new file mode 100644
index 00000000..a2fe13b6
--- /dev/null
+++ b/.github/workflows/probe-self-hosted-runners.yml
@@ -0,0 +1,113 @@
+# Manual probe: can this repo schedule a job on each NPU runner label,
+# pull the matching AscendHub CANN image, and see that many cards?
+#
+# Label shape: linux-aarch64-<chip>-<npu>
+#   chip: a2 (910B) or a3 (910C)
+#   npu:  available card count for that scale set (1 / 2 / 4 / 8 / 16)
+#
+# How to read a run:
+#   - Job stays queued: this repo has no matching scale set, or no idle capacity.
+#   - Job starts then fails on image pull / container create: runner is up; the
+#     container path is broken (wrong tag, hook, or device plugin).
+#   - Job is green and npu-smi lists the expected number of devices: that
+#     label is usable.
+#
+# schedule stays commented. Do not enable until someone decides this should
+# be a standing health check.
+name: Probe self-hosted runners
+
+on:
+  workflow_dispatch:
+  # schedule:
+  #   - cron: '17 */6 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: true
+
+permissions:
+  contents: read
+
+jobs:
+  probe:
+    name: linux-aarch64-${{ matrix.chip }}-${{ matrix.npu }}
+    runs-on: linux-aarch64-${{ matrix.chip }}-${{ matrix.npu }}
+    timeout-minutes: 20
+    strategy:
+      fail-fast: false
+      matrix:
+        chip: [a2, a3]
+        npu: [1, 2, 4, 8, 16]
+    defaults:
+      run:
+        shell: bash
+    # container.image is required on these ARC runners; without it the job
+    # may start but NPU is not allocated. Do not use ${{ env.* }} here —
+    # that context is not available for image.
+    container:
+      image: ${{ matrix.chip == 'a2' && 'swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12' || 'swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-a3-ubuntu22.04-py3.12' }}
+    steps:
+      - name: Probe host and NPU
+        env:
+          RUNNER_LABEL: linux-aarch64-${{ matrix.chip }}-${{ matrix.npu }}
+          IMAGE: ${{ matrix.chip == 'a2' && 'swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-910b-ubuntu22.04-py3.12' || 'swr.cn-south-1.myhuaweicloud.com/ascendhub/cann:9.1.0-a3-ubuntu22.04-py3.12' }}
+          NPU_COUNT: ${{ matrix.npu }}
+        run: |
+          set -euo pipefail
+          export PATH="/usr/local/sbin:/usr/local/bin:${PATH}"
+          if [ -f /usr/local/Ascend/ascend-toolkit/set_env.sh ]; then
+            # shellcheck disable=SC1091
+            source /usr/local/Ascend/ascend-toolkit/set_env.sh
+          fi
+
+          ids=()
+          i=0
+          while [ "${i}" -lt "${NPU_COUNT}" ]; do
+            ids+=("${i}")
+            i=$((i + 1))
+          done
+          IFS=,
+          export ASCEND_RT_VISIBLE_DEVICES="${ids[*]}"
+          unset IFS
+
+          echo "=== runner ==="
+          echo "RUNNER_NAME=${RUNNER_NAME:-}"
+          echo "RUNNER_OS=${RUNNER_OS:-}"
+          echo "RUNNER_ARCH=${RUNNER_ARCH:-}"
+          echo "RUNNER_LABEL=${RUNNER_LABEL}"
+          echo "IMAGE=${IMAGE}"
+          echo "NPU_COUNT=${NPU_COUNT}"
+          echo "ASCEND_RT_VISIBLE_DEVICES=${ASCEND_RT_VISIBLE_DEVICES}"
+          hostname || true
+          uname -a
+
+          echo "=== devices ==="
+          ls -l /dev/davinci* /dev/davinci_manager /dev/devmm_svm /dev/hisi_hdc 2>&1 || true
+
+          echo "=== npu-smi ==="
+          npu-smi info
+
+          npu_nodes="$(find /dev -maxdepth 1 -type c -name 'davinci[0-9]*' | wc -l | tr -d ' ')"
+          echo "davinci_nodes=${npu_nodes}"
+          if [ "${npu_nodes}" -lt "${NPU_COUNT}" ]; then
+            echo "expected at least ${NPU_COUNT} NPU device nodes on ${RUNNER_LABEL}, got ${npu_nodes}" >&2
+            exit 1
+          fi
+
+  summarize:
+    name: summarize
+    needs: probe
+    if: always()
+    runs-on: ubuntu-latest
+    steps:
+      - name: Report probe result
+        env:
+          PROBE_RESULT: ${{ needs.probe.result }}
+        run: |
+          set -euo pipefail
+          echo "probe matrix result: ${PROBE_RESULT}"
+          if [ "${PROBE_RESULT}" != "success" ]; then
+            echo "one or more runner labels failed or did not finish" >&2
+            exit 1
+          fi
+          echo "all linux-aarch64-a2/a3 labels (1/2/4/8/16 cards) picked up a job and saw NPU"