ci(musa): add S5000 CI workflow

Author: HermiaHuan

.github/configs/musa.yml

@@ -0,0 +1,23 @@
+# Copyright 2026 FlagOS Contributors
+# Licensed under the Apache License, Version 2.0.
+
+platform: musa
+
+ci_image: harbor.baai.ac.cn/flagos-dev/vllm-plugin-fl:29ceebd-musa-ci
+
+runner_labels:
+  - mt-cicd-vllm-plugin
+
+container_volumes:
+  - /data:/data
+  - /workspace:/workspace
+  - /hsdatapool:/hsdatapool
+
+container_options: >-
+  --hostname vllm-plugin-fl
+  --privileged
+  --ipc=host
+  --shm-size=64g
+  --env GEMS_VENDOR=mthreads
+  --env VLLM_PLUGINS=fl
+  --env MTHREADS_VISIBLE_DEVICES=all

.github/configs/platforms.yml

@@ -26,3 +26,5 @@ platforms:
     enabled: false
   hygon:
     enabled: true
+  musa:
+    enabled: true

.github/scripts/musa/check.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# Copyright (c) 2026 BAAI. All rights reserved.
+# Check Moore Threads MUSA availability.
+set -euo pipefail
+
+echo "Current time: $(date '+%Y-%m-%d %H:%M:%S')"
+echo "=== Checking Moore Threads MUSA availability ==="
+
+if command -v mthreads-gmi >/dev/null 2>&1; then
+  mthreads-gmi
+else
+  echo "::warning::mthreads-gmi not found; checking through torch_musa."
+fi
+
+python - <<'PY'
+import torch
+import torch_musa
+
+assert torch.musa.is_available(), "MUSA accelerator is unavailable"
+count = torch.musa.device_count()
+assert count > 0, "No MUSA devices detected"
+
+tensor = torch.ones((32, 32), device="musa:0")
+torch.musa.synchronize()
+
+print(f"MUSA devices: {count}")
+print(f"Tensor smoke: {tensor.device} {tuple(tensor.shape)}")
+PY

.github/scripts/musa/setup.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+# Copyright (c) 2026 BAAI. All rights reserved.
+# Setup script for Moore Threads MUSA CI environment.
+set -euo pipefail
+
+git config --global --add safe.directory "$(pwd)"
+
+: "${GEMS_VENDOR:?GEMS_VENDOR is not set}"
+: "${VLLM_PLUGINS:?VLLM_PLUGINS is not set}"
+: "${MTHREADS_VISIBLE_DEVICES:?MTHREADS_VISIBLE_DEVICES is not set}"
+
+python -m pip install --no-build-isolation --no-deps -e .
+
+python - <<'PY'
+import flag_gems
+import torch
+import torch_musa
+import vllm
+import vllm_fl
+from vllm.platforms import current_platform
+
+assert torch.musa.is_available(), "MUSA accelerator is unavailable"
+assert torch.musa.device_count() > 0, "No MUSA devices detected"
+assert current_platform.device_type == "musa", current_platform.device_type
+
+print(f"vLLM import ok: {vllm.__version__}")
+print(f"vLLM-FL import ok: {vllm_fl.__file__}")
+print(f"FlagGems import ok: {getattr(flag_gems, '__version__', 'unknown')}")
+print(f"Torch import ok: {torch.__version__}")
+print(f"MUSA available: {torch.musa.is_available()}")
+print(f"MUSA devices: {torch.musa.device_count()}")
+print(f"Platform: {current_platform}")
+PY

.github/workflows/_benchmark_test.yml

@@ -90,6 +90,7 @@ jobs:
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: benchmark-${{ inputs.platform }}
           path: |

.github/workflows/_build_wheel.yml

@@ -77,6 +77,7 @@ jobs:
       - name: Upload wheel
         if: inputs.upload-artifact
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: wheel
           path: dist/*.whl

.github/workflows/_e2e_test.yml

@@ -108,6 +108,7 @@ jobs:
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: e2e-${{ inputs.platform }}-${{ inputs.device }}-${{ inputs.task }}
           path: |

.github/workflows/_functional_test.yml

@@ -90,6 +90,7 @@ jobs:
       - name: Upload test results
         if: always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: functional-${{ inputs.platform }}
           path: |

.github/workflows/_unit_test.yml

@@ -89,6 +89,7 @@ jobs:
       - name: Upload coverage report
         if: inputs.upload_coverage && always()
         uses: actions/upload-artifact@v4
+        continue-on-error: true
         with:
           name: coverage-${{ inputs.platform }}
           path: |

.github/workflows/ci.yml

@@ -133,3 +133,14 @@ jobs:
     with:
       platform: hygon
     secrets: inherit
+
+  # ============================================================
+  # Job 4d: MUSA platform testing
+  # ============================================================
+  test-musa:
+    needs: discover
+    if: contains(fromJson(needs.discover.outputs.platforms), 'musa')
+    uses: ./.github/workflows/_platform_test.yml
+    with:
+      platform: musa
+    secrets: inherit