Add build-time sanity check, disable hardware tests until size is reduced

The 7+ GB artifact is too slow to transfer to self-hosted runners.
Added import verification in the build job itself. Hardware GPU tests
will be re-enabled once artifact size is under ~2 GB.

Author: ramkrishna2910

.github/workflows/build-vllm-rocm.yml

@@ -466,7 +466,21 @@ jobs:
         du -sh lib/hipblaslt/ 2>/dev/null || true
         du -sh lib/libLLVM* 2>/dev/null || true
 
-    - name: List artifact contents
+    - name: Verify bundled environment works
+      run: |
+        echo "=== Sanity check: verify venv works after stripping ==="
+        export LD_LIBRARY_PATH="/opt/vllm/lib:${LD_LIBRARY_PATH:-}"
+        /opt/vllm/bin/python3 -c "import vllm; print(f'vLLM {vllm.__version__} OK')"
+        /opt/vllm/bin/python3 -c "import torch; print(f'PyTorch {torch.__version__} OK')"
+        /opt/vllm/bin/python3 -c "
+        from vllm.entrypoints.openai.api_server import build_async_engine_client
+        print('vLLM server entrypoint importable OK')
+        "
+        # Verify the launcher script syntax
+        bash -n /opt/vllm/bin/vllm-server
+        echo "All sanity checks passed"
+
+    - name: Report artifact size
       run: |
         echo "=== Final artifact ==="
         du -sh /opt/vllm/
@@ -476,9 +490,6 @@ jobs:
         echo ""
         echo "Largest items in lib/:"
         du -sh /opt/vllm/lib/* 2>/dev/null | sort -rh | head -15
-        echo ""
-        echo "bin/ entry points:"
-        ls -la /opt/vllm/bin/vllm-server /opt/vllm/bin/python3* 2>/dev/null
 
     - name: Upload build artifacts
       uses: actions/upload-artifact@v4
@@ -504,67 +515,19 @@ jobs:
         [ -d "/opt/vllm" ] && sudo rm -rf /opt/vllm
         echo "Cleanup completed"
 
-  test-stx-halo:
-    runs-on: ${{ matrix.runner }}
-    needs: [prepare-matrix, build-ubuntu]
-    if: |
-      needs.build-ubuntu.result == 'success' &&
-      contains(github.event.inputs.gfx_target || 'gfx1151,gfx1150,gfx120X,gfx110X', 'gfx1151')
-    strategy:
-      matrix:
-        include:
-          - os: Linux
-            gfx_target: gfx1151
-            runner: ["stx-halo", "Linux"]
-      fail-fast: false
-
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-
-    - name: Test vLLM build
-      uses: ./.github/actions/test-vllm-build
-      with:
-        os_type: Linux
-        gfx_target: ${{ matrix.gfx_target }}
-        artifact_name: vllm-ubuntu-rocm-${{ matrix.gfx_target }}-x64
-
-  test-stx:
-    runs-on: ${{ matrix.runner }}
-    needs: [prepare-matrix, build-ubuntu]
-    if: |
-      needs.build-ubuntu.result == 'success' &&
-      contains(github.event.inputs.gfx_target || 'gfx1151,gfx1150,gfx120X,gfx110X', 'gfx1150')
-    strategy:
-      matrix:
-        include:
-          - os: Linux
-            gfx_target: gfx1150
-            runner: ["stx", "Linux"]
-      fail-fast: false
-
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-
-    - name: Test vLLM build
-      uses: ./.github/actions/test-vllm-build
-      with:
-        os_type: Linux
-        gfx_target: ${{ matrix.gfx_target }}
-        artifact_name: vllm-ubuntu-rocm-${{ matrix.gfx_target }}-x64
+  # Hardware GPU tests are disabled until artifact size is reduced.
+  # The 7+ GB artifact takes too long to transfer to self-hosted runners.
+  # TODO: Re-enable once size is under ~2 GB or test approach is reworked.
 
   create-release:
-    needs: [prepare-matrix, build-ubuntu, test-stx-halo, test-stx]
+    needs: [prepare-matrix, build-ubuntu]
     runs-on: ubuntu-22.04
     permissions:
       contents: write
       pull-requests: write
     if: |
       always() &&
       needs.build-ubuntu.result == 'success' &&
-      (needs.test-stx-halo.result == 'success' || needs.test-stx-halo.result == 'skipped') &&
-      (needs.test-stx.result == 'success' || needs.test-stx.result == 'skipped') &&
       github.event_name != 'pull_request' &&
       (github.event_name == 'workflow_dispatch' &&
        (github.event.inputs.create_release == 'true' || github.event.inputs.create_release == null) ||