fix(xpu): stabilize XPU test workers

lib/gpu_memory_service/tests/test_v1_mempool_client.py

@@ -5,17 +5,19 @@
 from unittest.mock import Mock
 
 import pytest
-from _deps import HAS_GMS
+from _deps import HAS_GMS, HAS_TORCH
 
 if not HAS_GMS:
     pytest.skip(
         "gpu_memory_service package is not available in this test image",
         allow_module_level=True,
     )
 
-torch = pytest.importorskip("torch", reason="torch is required")
+if not HAS_TORCH:
+    pytest.skip("PyTorch is required", allow_module_level=True)
 
 import gpu_memory_service.v1.client.mempool as mempool_module  # noqa: E402
+import torch
 from gpu_memory_service.v1.client.mempool import TorchMempoolMemoryClient  # noqa: E402
 
 pytestmark = [

tests/fault_tolerance/etcd_ha/test_vllm.py

@@ -23,6 +23,7 @@
 from tests.utils.constants import FAULT_TOLERANCE_MODEL_NAME, DynamoPortRange
 from tests.utils.device import (
     build_nixl_kv_transfer_config,
+    detect_target_device,
     get_default_vllm_block_size,
 )
 from tests.utils.engine_process import FRONTEND_PORT
@@ -96,6 +97,13 @@ def __init__(
         env["ETCD_ENDPOINTS"] = ",".join(etcd_endpoints)
         env["DYN_SYSTEM_USE_ENDPOINT_HEALTH_STATUS"] = '["generate"]'
         env["DYN_SYSTEM_PORT"] = port
+        if detect_target_device() == "xpu":
+            visible_devices = [
+                device.strip()
+                for device in env.get("ZE_AFFINITY_MASK", "").split(",")
+                if device.strip()
+            ]
+            env["ZE_AFFINITY_MASK"] = visible_devices[0] if visible_devices else "0"
 
         # Both prefill and decode workers need kv-transfer-config for disaggregated mode
         if mode != WorkerMode.AGGREGATED:

tests/serve/test_vllm_omni.py

@@ -23,6 +23,7 @@
     params_with_model_mark,
     run_serve_deployment,
 )
+from tests.utils.device import detect_target_device
 from tests.utils.engine_process import EngineConfig
 from tests.utils.payloads import (
     AudioSpeechPayload,
@@ -261,4 +262,11 @@ def test_omni_serve_deployment(
     config = dataclasses.replace(
         vllm_omni_config_test, frontend_port=dynamo_dynamic_ports.frontend_port
     )
-    run_serve_deployment(config, request, ports=dynamo_dynamic_ports)
+    extra_env = (
+        {"_PROFILE_OVERRIDE_VLLM_KV_CACHE_BYTES": "536870912"}
+        if config.name == "omni_audio" and detect_target_device() == "xpu"
+        else None
+    )
+    run_serve_deployment(
+        config, request, ports=dynamo_dynamic_ports, extra_env=extra_env
+    )

tests/serve/test_vllm_xpu.py

@@ -65,6 +65,7 @@ class VLLMConfig(EngineConfig):
 def clear_stale_fpm_env(monkeypatch):
     """Keep serve/XPU tests isolated from stale FPM env vars inherited by CI or parent shells."""
     monkeypatch.delenv("DYN_FORWARDPASS_METRIC_PORT", raising=False)
+    monkeypatch.delenv("DYN_FPM_TRACE", raising=False)
 
 
 # Generated multimodal configs from profile definitions