Update XPU Docker runtime stack & hf_home config (#23820)

Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>

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

@@ -72,8 +72,6 @@ jobs:
     needs: [check-changes, pr-gate]
     if: needs.check-changes.outputs.main_package == 'true'
     runs-on: intel-bmg
-    env:
-      HF_HOME: /home/sdp/.cache/huggingface
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -99,8 +97,10 @@ jobs:
           container_id=$(docker run -dt \
             --group-add 992 \
             --group-add $(getent group video | cut -d: -f3) \
-            -v ${HF_HOME}:/root/.cache/huggingface \
+            --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"

docker/xpu.Dockerfile

@@ -20,6 +20,16 @@ ARG SG_LANG_KERNEL_BRANCH=main
 RUN useradd -m -d /home/sdp -s /bin/bash sdp && \
     chown -R sdp:sdp /home/sdp
 
+USER root
+
+# Install the latest UMD driver for SYCL-TLA
+RUN apt-get install -y software-properties-common && \
+    add-apt-repository -y ppa:kobuk-team/intel-graphics && \
+    apt-get update  && \
+    apt-get install -y libze-intel-gpu1 libze1 intel-metrics-discovery intel-opencl-icd clinfo intel-gsc && \
+    apt-get install -y intel-media-va-driver-non-free libmfx-gen1 libvpl2 libvpl-tools libva-glx2 va-driver-all vainfo && \
+    apt-get install -y libze-dev intel-ocloc
+
 # Switch to non-root user 'sdp'
 USER sdp
 
@@ -38,12 +48,6 @@ RUN curl -fsSL -v -o miniforge.sh -O https://github.com/conda-forge/miniforge/re
     # Append environment activation to .bashrc for interactive shells
     echo ". /home/sdp/miniforge3/bin/activate; conda activate py${PYTHON_VERSION}; . /opt/intel/oneapi/setvars.sh; cd /home/sdp" >> /home/sdp/.bashrc
 
-USER root
-RUN apt-get update && apt install -y intel-ocloc
-
-# Switch back to user sdp
-USER sdp
-
 RUN --mount=type=secret,id=github_token \
     cd /home/sdp && \
     . /home/sdp/miniforge3/bin/activate && \

test/srt/xpu/test_deepseek_ocr.py

@@ -2,7 +2,6 @@
 python3 -m unittest test_deepseek_ocr.py
 """
 
-import gc
 import json
 import os
 import unittest
@@ -21,22 +20,8 @@ from sglang.test.test_utils import (
 
 
 class TestDeepSeekOCR(CustomTestCase):
-    @classmethod
-    def _cleanup_xpu_memory(cls):
-        gc.collect()
-        try:
-            import torch
-
-            if hasattr(torch, "xpu") and torch.xpu.is_available():
-                torch.xpu.synchronize()
-                torch.xpu.empty_cache()
-        except Exception:
-            # Best-effort cleanup only; tests should continue if cleanup is unavailable.
-            pass
-
     @classmethod
     def setUpClass(cls):
-        cls._cleanup_xpu_memory()
         cls.model = "deepseek-ai/DeepSeek-OCR"
         cls.tokenizer = get_tokenizer(cls.model)
         cls.base_url = DEFAULT_URL_FOR_TEST
@@ -65,8 +50,12 @@ class TestDeepSeekOCR(CustomTestCase):
     def tearDownClass(cls):
         """Fixture that is run once after all tests in the class."""
         if hasattr(cls, "process") and cls.process:
+            cls.process.terminate()
+            try:
+                cls.process.wait(timeout=30)
+            except Exception:
+                # Force kill if it didn't exit cleanly in time
                 kill_process_tree(cls.process.pid)
-        cls._cleanup_xpu_memory()
 
     def get_request_json(self, max_new_tokens=32, n=1):
         response = requests.post(

test/srt/xpu/test_deepseek_ocr_triton.py

@@ -6,7 +6,7 @@ import os
 import unittest
 from pathlib import Path
 
-import test_deepseek_ocr as deepseek_ocr
+from test_deepseek_ocr import TestDeepSeekOCR
 
 from sglang.srt.utils.hf_transformers import get_tokenizer
 from sglang.test.test_utils import (
@@ -16,10 +16,11 @@ from sglang.test.test_utils import (
 )
 
 
-class TestDeepSeekOCRTriton(deepseek_ocr.TestDeepSeekOCR):
+# TODO: Temporarily disable this test and re-enable it after Triton-XPU is upgraded.
+@unittest.skip("Temporarily disabled until Triton-XPU upgrade")
+class TestDeepSeekOCRTriton(TestDeepSeekOCR):
     @classmethod
     def setUpClass(cls):
-        cls._cleanup_xpu_memory()
         cls.model = "deepseek-ai/DeepSeek-OCR"
         cls.tokenizer = get_tokenizer(cls.model)
         cls.base_url = DEFAULT_URL_FOR_TEST
@@ -45,5 +46,8 @@ class TestDeepSeekOCRTriton(deepseek_ocr.TestDeepSeekOCR):
         )
 
 
+# Prevent pytest from collecting the imported base test class here.
+del TestDeepSeekOCR
+
 if __name__ == "__main__":
     unittest.main()

test/srt/xpu/test_intel_xpu_backend.py

@@ -3,7 +3,6 @@ Usage:
 python3 -m unittest test_intel_xpu_backend.TestIntelXPUBackend.test_latency_qwen_model
 """
 
-import gc
 import unittest
 from functools import wraps
 
@@ -16,29 +15,17 @@ from sglang.test.test_utils import (
 )
 
 
-def _cleanup_xpu_memory():
+def intel_xpu_benchmark(
-    gc.collect()
+    extra_args=None, min_throughput=None, mem_fraction_static="0.4"
-    try:
+):
-        import torch
-
-        if hasattr(torch, "xpu") and torch.xpu.is_available():
-            torch.xpu.synchronize()
-            torch.xpu.empty_cache()
-    except Exception:
-        # Best-effort cleanup only.
-        pass
-
-
-def intel_xpu_benchmark(extra_args=None, min_throughput=None):
     def decorator(test_func):
         @wraps(test_func)
         def wrapper(self):
-            _cleanup_xpu_memory()
             common_args = [
                 "--disable-radix",
                 "--trust-remote-code",
                 "--mem-fraction-static",
-                "0.4",
+                str(mem_fraction_static),
                 "--batch-size",
                 "1",
                 "--device",
@@ -48,12 +35,9 @@ def intel_xpu_benchmark(extra_args=None, min_throughput=None):
             full_args = common_args + ci_args + (extra_args or [])
 
             model = test_func(self)
-            try:
+            prefill_latency, decode_throughput, decode_latency = run_bench_one_batch(
-                prefill_latency, decode_throughput, decode_latency = (
+                model, full_args
-                    run_bench_one_batch(model, full_args)
             )
-            finally:
-                _cleanup_xpu_memory()
 
             print(f"{model=}")
             print(f"{prefill_latency=}")
@@ -70,11 +54,14 @@ def intel_xpu_benchmark(extra_args=None, min_throughput=None):
 
 class TestIntelXPUBackend(CustomTestCase):
 
-    @intel_xpu_benchmark(min_throughput=10)
+    @intel_xpu_benchmark(min_throughput=10, mem_fraction_static="0.3")
     def test_latency_qwen_model(self):
         return DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN
 
-    @intel_xpu_benchmark(["--attention-backend", "intel_xpu", "--page-size", "128"])
+    @intel_xpu_benchmark(
+        ["--attention-backend", "intel_xpu", "--page-size", "128"],
+        mem_fraction_static="0.5",
+    )
     def test_attention_backend(self):
         return DEFAULT_SMALL_MODEL_NAME_FOR_TEST_BASE