[ci] xpu: trigger pr-test-xpu on multimodal_gen changes

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

@@ -28,7 +28,9 @@ jobs:
   check-changes:
     runs-on: ubuntu-latest
     outputs:
-      main_package: ${{ steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }}
+      changes_exist: ${{ steps.filter.outputs.main_package == 'true' || steps.filter.outputs.multimodal_gen == 'true' || steps.run-mode.outputs.run_all_tests == 'true' }}
+      main_package: ${{ steps.filter.outputs.main_package == 'true' || steps.run-mode.outputs.run_all_tests == 'true' }}
+      multimodal_gen: ${{ steps.filter.outputs.multimodal_gen == 'true' || steps.run-mode.outputs.run_all_tests == 'true' }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -60,11 +62,16 @@ jobs:
               - "python/sglang/kernels/aot/**/!(*.md|THIRDPARTYNOTICES.txt|LICENSE)"
               - ".github/workflows/pr-test-xpu.yml"
               - "docker/xpu.Dockerfile"
+            multimodal_gen:
+              - "python/sglang/multimodal_gen/**/!(*.md|*.ipynb)"
+              - "python/pyproject_xpu.toml"
+              - ".github/workflows/pr-test-xpu.yml"
+              - "docker/xpu.Dockerfile"
 
   # ==================== PR Gate ==================== #
   pr-gate:
     needs: check-changes
-    if: needs.check-changes.outputs.main_package == 'true'
+    if: needs.check-changes.outputs.changes_exist == 'true'
     uses: ./.github/workflows/pr-gate.yml
     secrets: inherit
 
@@ -249,15 +256,96 @@ jobs:
             docker rmi -f "${CI_SGLANG_XPU_IMAGE}" || true
           fi
 
+  # ==================== Multimodal Gen ==================== #
+  multimodal-gen-test-1-gpu-xpu:
+    needs: [check-changes, pr-gate]
+    if: needs.check-changes.outputs.multimodal_gen == 'true'
+    runs-on: bmg-multigen-models
+    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 "lmcache>=0.3.9"
+          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
+          # Fetch tags so setuptools_scm resolves a real version instead of
+          # falling back to 0.0.0 on a shallow/tag-less checkout.
+          docker exec -w /sglang-checkout ci_sglang_xpu git fetch origin '+refs/tags/*:refs/tags/*' --force
+          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 /bin/bash -c '/opt/venv/bin/hf auth login --token ${HF_TOKEN}'
+
+      - name: Run diffusion server tests (1-GPU)
+        timeout-minutes: 60
+        run: |
+          docker exec ci_sglang_xpu bash -c "source /opt/venv/bin/activate && cd /sglang-checkout/python && python3 sglang/multimodal_gen/test/run_suite.py --suite 1-gpu-xpu"
+
+      - name: Cleanup container
+        if: always()
+        run: |
+          docker run --rm -v "${{ github.workspace }}:/w" busybox:latest \
+            chown -R "$(id -u):$(id -g)" /w || true
+          rm -rf \
+            python/build \
+            python/dist \
+            python/sglang.egg-info \
+            python/sglang/*.egg-info \
+            test/result.jsonl \
+            test/results \
+            test/.pytest_cache \
+            .pytest_cache || true
+          find . -type d -name "__pycache__" -prune -exec rm -rf {} + || true
+          find . -type f -name "*.pyc" -delete || true
+          # SIGTERM sglang and drain GPU context before `docker rm -f`;
+          # SIGKILL leaves the xe/GuC exec queue registered and triggers a
+          # GT reset (+ devcoredump) on B580.
+          if docker ps --format '{{.Names}}' | grep -qx ci_sglang_xpu; then
+            docker exec ci_sglang_xpu bash -c '
+              pkill -TERM -f "sglang|run_suite|python3.*test_" 2>/dev/null || true
+              for _ in $(seq 1 30); do
+                pgrep -f "sglang::|sglang.launch_server" >/dev/null || break
+                sleep 1
+              done
+              pkill -KILL -f "sglang|run_suite" 2>/dev/null || true
+            ' || true
+          fi
+          docker rm -f ci_sglang_xpu || true
+          if [[ -n "${CI_SGLANG_XPU_IMAGE:-}" ]]; then
+            docker rmi -f "${CI_SGLANG_XPU_IMAGE}" || true
+          fi
+
   finish:
     if: always()
-    needs: [stage-a-test-1-gpu-xpu, stage-b-test-1-gpu-xpu, pr-gate]
+    needs: [stage-a-test-1-gpu-xpu, stage-b-test-1-gpu-xpu, multimodal-gen-test-1-gpu-xpu, pr-gate]
     runs-on: ubuntu-latest
     steps:
       - name: Check job status
         run: |
           stage_a="${{ needs.stage-a-test-1-gpu-xpu.result }}"
           stage_b="${{ needs.stage-b-test-1-gpu-xpu.result }}"
+          multimodal_gen="${{ needs.multimodal-gen-test-1-gpu-xpu.result }}"
           if [ "$stage_a" != "success" ] && [ "$stage_a" != "skipped" ]; then
             echo "stage-a failed with result: $stage_a"
             exit 1
@@ -266,5 +354,9 @@ jobs:
             echo "stage-b failed with result: $stage_b"
             exit 1
           fi
+          if [ "$multimodal_gen" != "success" ] && [ "$multimodal_gen" != "skipped" ]; then
+            echo "multimodal-gen failed with result: $multimodal_gen"
+            exit 1
+          fi
           echo "All jobs completed successfully"
           exit 0

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

@@ -1223,8 +1223,11 @@ def _prepare_denoising_loop(self, batch: Req, server_args: ServerArgs):
         image_kwargs = self.prepare_extra_func_kwargs(
             getattr(self.transformer, "forward", self.transformer),
             {
+                # Pass None (not []) so T2V paths whose transformer has no
+                # image_embedder skip the branch; diffusers guards on
+                # `is not None` only.
                 # TODO: make sure on-device
-                "encoder_hidden_states_image": image_embeds,
+                "encoder_hidden_states_image": image_embeds if image_embeds else None,
             },
         )
 

python/sglang/multimodal_gen/test/server/gpu_cases.py

@@ -1233,6 +1233,40 @@ def _make_5090_h3_consumer_budget_case() -> DiffusionTestCase:
 ONE_GPU_5090_CASES.append(_make_5090_h3_consumer_budget_case())
 
 
+# Intel Arc Pro B60 has 24 GiB of XPU memory, so only sub-~5B-parameter
+# checkpoints fit fully resident. Larger cases in ONE_GPU_CASES (FLUX.1-dev,
+# FLUX.2-dev, Qwen-Image, Hunyuan3D, SANA-Video, image-edit families) OOM on
+# 24 GiB, and the FP8/NVFP4 quant paths are CUDA-only.
+ONE_GPU_XPU_CASE_IDS = (
+    "zimage_image_t2i",
+    "flux_2_klein_image_t2i",
+    "flux_2_klein_base_image_t2i",
+    "wan2_1_t2v_1.3b",
+)
+
+
+def _select_xpu_cases(case_ids: tuple[str, ...]) -> list[DiffusionTestCase]:
+    cases_by_id = {case.id: case for case in ONE_GPU_CASES}
+    missing = [case_id for case_id in case_ids if case_id not in cases_by_id]
+    if missing:
+        raise RuntimeError(f"Unknown XPU diffusion case(s): {missing}")
+    return [cases_by_id[case_id] for case_id in case_ids]
+
+
+# Consistency GT images are H100-generated; XPU output diverges at the
+# pixel level (different attention kernels + fp reductions on Xe2) so
+# SSIM/PSNR against the H100 golden always fails. test_server_1_gpu.py
+# parametrizes directly from ONE_GPU_CASES, so mutate those entries in
+# place -- overriding only via ONE_GPU_XPU_CASES would be ignored.
+if current_platform.is_xpu():
+    _xpu_ids = set(ONE_GPU_XPU_CASE_IDS)
+    for _i, _case in enumerate(ONE_GPU_CASES):
+        if _case.id in _xpu_ids and _case.run_consistency_check:
+            ONE_GPU_CASES[_i] = replace(_case, run_consistency_check=False)
+
+ONE_GPU_XPU_CASES = _select_xpu_cases(ONE_GPU_XPU_CASE_IDS)
+
+
 # Nested unit/ tests verified to pass on AMD/ROCm as-is (no code change).
 # Enabled incrementally and AMD-only: the CUDA `multimodal-gen-unit-test`
 # lane keeps the flat glob below. Files that still need fixes/skips are added
@@ -1300,6 +1334,9 @@ def _discover_unit_tests() -> list[str]:
     "1-gpu-5090": [
         ("test_server_1_gpu_5090.py", ONE_GPU_5090_CASES),
     ],
+    "1-gpu-xpu": [
+        ("test_server_1_gpu.py", ONE_GPU_XPU_CASES),
+    ],
     "2-gpu": [
         ("test_server_2_gpu.py", TWO_GPU_CASES),
     ],