[XPU CI] Re-enable stage B with docker-pull flow and split tests (#27526)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.github/workflows/pr-test-xpu.yml

@@ -127,74 +127,70 @@ jobs:
     steps:
       - run: echo "stage-a passed"
 
-  # ==================== Stage B (disabled — tests moved to Stage A) ==================== #
-  # stage-b-test-1-gpu-xpu:
-  #   needs: [check-changes, pr-gate, wait-for-stage-a]
-  #   if: needs.check-changes.outputs.main_package == 'true'
-  #   runs-on: intel-bmg
-  #   steps:
-  #     - name: Checkout code
-  #       uses: actions/checkout@v4
-  #       with:
-  #         fetch-depth: 0
-  #         ref: ${{ inputs.ref || github.ref }}
-  #
-  #     - name: Build Docker image
-  #       run: |
-  #         PR_REPO=${{ github.event.pull_request.head.repo.clone_url }}
-  #         PR_HEAD_REF=${{ github.head_ref }}
-  #         docker build \
-  #           ${PR_REPO:+--build-arg SG_LANG_REPO=$PR_REPO} \
-  #           ${PR_HEAD_REF:+--build-arg SG_LANG_BRANCH=$PR_HEAD_REF} \
-  #           --no-cache --progress=plain -f docker/xpu.Dockerfile -t xpu_sglang_main:bmg .
-  #
-  #     - name: Run container
-  #       id: start_container
-  #       run: |
-  #         container_id=$(docker run -dt \
-  #           --group-add 992 \
-  #           --group-add $(getent group video | cut -d: -f3) \
-  #           --group-add $(getent group render | cut -d: -f3) \
-  #           -v $HOME/.cache/huggingface:/root/.cache/huggingface \
-  #           --device /dev/dri \
-  #           -v /dev/dri/by-path:/dev/dri/by-path \
-  #           -e HF_TOKEN="$(cat ~/huggingface_token.txt)" \
-  #           xpu_sglang_main:bmg)
-  #         echo "Started container: $container_id"
-  #         echo "container_id=$container_id" >> "$GITHUB_OUTPUT"
-  #
-  #     - name: Install Dependency
-  #       timeout-minutes: 20
-  #       run: |
-  #         cid="${{ steps.start_container.outputs.container_id }}"
-  #         docker exec "$cid" /opt/venv/bin/python3 -m pip install --upgrade pip
-  #         docker exec "$cid" /opt/venv/bin/python3 -m pip install pytest expecttest ray huggingface_hub tabulate
-  #         docker exec "$cid" /opt/venv/bin/python3 -m pip uninstall -y flashinfer-python
-  #         docker exec "$cid" /bin/bash -c '/opt/venv/bin/hf auth login --token ${HF_TOKEN} '
-  #
-  #     - name: Run stage-b tests
-  #       timeout-minutes: 60
-  #       run: |
-  #         cid="${{ steps.start_container.outputs.container_id }}"
-  #         docker exec "$cid" bash -c "source /opt/venv/bin/activate && cd /sgl-workspace/sglang/test && python3 run_suite.py --hw xpu --suite stage-b-test-1-gpu-xpu"
-  #
-  #     - name: Cleanup container
-  #       if: always()
-  #       run: |
-  #         cid="${{ steps.start_container.outputs.container_id }}"
-  #         docker rm -f "$cid" || true
+  # ==================== Stage B ==================== #
+  stage-b-test-1-gpu-xpu:
+    needs: [check-changes, pr-gate, wait-for-stage-a]
+    if: needs.check-changes.outputs.main_package == 'true'
+    runs-on: intel-bmg
+    env:
+      DOCKERHUB_INTEL_USERNAME: ${{ secrets.DOCKERHUB_INTEL_USERNAME }}
+      DOCKERHUB_INTEL_TOKEN: ${{ secrets.DOCKERHUB_INTEL_TOKEN }}
+    steps:
+      - name: Reset workspace ownership
+        run: |
+          docker run --rm -v "${{ github.workspace }}:/w" busybox:latest \
+            chown -R "$(id -u):$(id -g)" /w || true
+
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ inputs.ref || github.ref }}
+
+      - name: Start CI container
+        run: |
+          export HF_TOKEN="$(cat ~/huggingface_token.txt)"
+          bash scripts/ci/xpu/xpu_ci_start_container.sh
+        env:
+          GITHUB_WORKSPACE: ${{ github.workspace }}
+
+      - name: Install Dependency
+        timeout-minutes: 60
+        run: |
+          docker exec ci_sglang_xpu /opt/venv/bin/python3 -m pip install --upgrade pip
+          docker exec ci_sglang_xpu /opt/venv/bin/python3 -m pip install pytest expecttest ray huggingface_hub tabulate
+          docker exec ci_sglang_xpu /opt/venv/bin/python3 -m pip uninstall -y flashinfer-python sgl-kernel sglang
+          docker exec ci_sglang_xpu cp /sglang-checkout/python/pyproject_xpu.toml /sglang-checkout/python/pyproject.toml
+          docker exec -w /sglang-checkout/python ci_sglang_xpu /opt/venv/bin/python3 -m pip install --no-cache-dir . --extra-index-url https://download.pytorch.org/whl/xpu
+          docker exec ci_sglang_xpu /opt/venv/bin/python3 -m pip install --no-cache-dir --no-deps xgrammar==0.1.33
+          docker exec ci_sglang_xpu /opt/venv/bin/python3 -m pip install triton-xpu==3.7.1 --index-url https://download.pytorch.org/whl/test/xpu --force-reinstall
+          docker exec ci_sglang_xpu /bin/bash -c '/opt/venv/bin/hf auth login --token ${HF_TOKEN}'
+
+      - name: Run stage-b tests
+        timeout-minutes: 60
+        run: |
+          docker exec ci_sglang_xpu bash -c "source /opt/venv/bin/activate && cd /sglang-checkout/test && python3 run_suite.py --hw xpu --suite stage-b-test-1-gpu-xpu"
+
+      - name: Cleanup container
+        if: always()
+        run: docker rm -f ci_sglang_xpu || true
 
   finish:
     if: always()
-    needs: [stage-a-test-1-gpu-xpu, pr-gate]
+    needs: [stage-a-test-1-gpu-xpu, stage-b-test-1-gpu-xpu, pr-gate]
     runs-on: ubuntu-latest
     steps:
       - name: Check job status
         run: |
           stage_a="${{ needs.stage-a-test-1-gpu-xpu.result }}"
+          stage_b="${{ needs.stage-b-test-1-gpu-xpu.result }}"
           if [ "$stage_a" != "success" ] && [ "$stage_a" != "skipped" ]; then
             echo "stage-a failed with result: $stage_a"
             exit 1
           fi
+          if [ "$stage_b" != "success" ] && [ "$stage_b" != "skipped" ]; then
+            echo "stage-b failed with result: $stage_b"
+            exit 1
+          fi
           echo "All jobs completed successfully"
           exit 0

test/registered/attention/test_chunk_gated_delta_rule.py

@@ -10,7 +10,7 @@ from sglang.srt.utils import get_device
 from sglang.test.ci.ci_register import register_cuda_ci, register_xpu_ci
 
 register_cuda_ci(est_time=11, stage="base-b", runner_config="1-gpu-large")
-register_xpu_ci(est_time=900, suite="stage-a-test-1-gpu-xpu")
+register_xpu_ci(est_time=900, suite="stage-b-test-1-gpu-xpu")
 
 
 @unittest.skipIf(

test/registered/xpu/test_deepseek_ocr.py

@@ -19,7 +19,7 @@ from sglang.test.test_utils import (
     popen_launch_server,
 )
 
-register_xpu_ci(est_time=360, suite="stage-a-test-1-gpu-xpu")
+register_xpu_ci(est_time=360, suite="stage-b-test-1-gpu-xpu")
 
 
 class TestDeepSeekOCR(CustomTestCase):

test/registered/xpu/test_deepseek_ocr_triton.py

@@ -18,7 +18,7 @@ from sglang.test.test_utils import (
 
 register_xpu_ci(
     est_time=360,
-    suite="stage-a-test-1-gpu-xpu",
+    suite="stage-b-test-1-gpu-xpu",
     disabled="Temporarily disabled until Triton-XPU upgrade",
 )
 

test/registered/xpu/test_gemma_4_e2b.py

@@ -130,7 +130,11 @@ class TestGemma4E2BXPU(CustomTestCase):
 from sglang.test.ci.ci_register import register_xpu_ci
 
 # Single e2e test: boot + a short Q&A.
-register_xpu_ci(est_time=240, suite="stage-b-test-1-gpu-xpu")
+register_xpu_ci(
+    est_time=240,
+    suite="stage-b-test-1-gpu-xpu",
+    disabled="OOM on stage-b XPU runner (server launch fails with --mem-fraction-static)",
+)
 
 if __name__ == "__main__":
     unittest.main()

test/registered/xpu/test_intel_xpu_backend.py

@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
     run_bench_one_batch,
 )
 
-register_xpu_ci(est_time=600, suite="stage-a-test-1-gpu-xpu")
+register_xpu_ci(est_time=600, suite="stage-b-test-1-gpu-xpu")
 
 
 def intel_xpu_benchmark(

test/registered/xpu/test_topk.py

@@ -11,7 +11,7 @@ from sglang.srt.layers.moe.topk import (
 from sglang.test.ci.ci_register import register_xpu_ci
 from sglang.test.test_utils import CustomTestCase
 
-register_xpu_ci(est_time=5, suite="stage-a-test-1-gpu-xpu")
+register_xpu_ci(est_time=5, suite="stage-b-test-1-gpu-xpu")
 
 
 # Nemotron-3 uses biased_grouped_topk