[pull] master from ray-project:master

doc/source/ray-core/scheduling/accelerators.rst

@@ -79,10 +79,11 @@ If you need to, you can :ref:`override <specify-node-resources>` this.
 
         .. tip::
 
-            You can set the ``ONEAPI_DEVICE_SELECTOR`` environment variable before starting a Ray node
+            You can set the ``ZE_AFFINITY_MASK`` environment variable before starting a Ray node
             to limit the Intel GPUs that are visible to Ray.
-            For example, ``ONEAPI_DEVICE_SELECTOR=1,3 ray start --head --num-gpus=2``
+            For example, ``ZE_AFFINITY_MASK=1,3 ray start --head --num-gpus=2``
             lets Ray only see devices 1 and 3.
+            ``ONEAPI_DEVICE_SELECTOR`` is still read as a fallback for backward compatibility.
 
     .. tab-item:: AWS Neuron Core
         :sync: AWS Neuron Core
@@ -282,12 +283,12 @@ and assign accelerators to the task or actor by setting the corresponding enviro
             class GPUActor:
                 def ping(self):
                     print("GPU IDs: {}".format(ray.get_runtime_context().get_accelerator_ids()["GPU"]))
-                    print("ONEAPI_DEVICE_SELECTOR: {}".format(os.environ["ONEAPI_DEVICE_SELECTOR"]))
+                    print("ZE_AFFINITY_MASK: {}".format(os.environ["ZE_AFFINITY_MASK"]))
 
             @ray.remote(num_gpus=1)
             def gpu_task():
                 print("GPU IDs: {}".format(ray.get_runtime_context().get_accelerator_ids()["GPU"]))
-                print("ONEAPI_DEVICE_SELECTOR: {}".format(os.environ["ONEAPI_DEVICE_SELECTOR"]))
+                print("ZE_AFFINITY_MASK: {}".format(os.environ["ZE_AFFINITY_MASK"]))
 
             gpu_actor = GPUActor.remote()
             ray.get(gpu_actor.ping.remote())
@@ -298,9 +299,9 @@ and assign accelerators to the task or actor by setting the corresponding enviro
             :options: +MOCK
 
             (GPUActor pid=52420) GPU IDs: [0]
-            (GPUActor pid=52420) ONEAPI_DEVICE_SELECTOR: 0
+            (GPUActor pid=52420) ZE_AFFINITY_MASK: 0
             (gpu_task pid=51830) GPU IDs: [1]
-            (gpu_task pid=51830) ONEAPI_DEVICE_SELECTOR: 1
+            (gpu_task pid=51830) ZE_AFFINITY_MASK: 1
 
     .. tab-item:: AWS Neuron Core
         :sync: AWS Neuron Core

doc/source/serve/llm/user-guides/sglang.md

@@ -25,6 +25,7 @@ Set the following environment variable before running any example:
 
 - **CUDA:** `RAY_EXPERIMENTAL_NOSET_CUDA_VISIBLE_DEVICES=0`
 - **ROCm:** `RAY_EXPERIMENTAL_NOSET_HIP_VISIBLE_DEVICES=0`
+- **Intel GPU:** `RAY_EXPERIMENTAL_NOSET_ZE_AFFINITY_MASK=0`
 
 ## Online serving (single node)
 

java/dependencies.bzl

@@ -6,7 +6,7 @@ def gen_java_deps():
         artifacts = [
             "com.fasterxml.jackson.core:jackson-databind:2.18.8",
             "com.github.java-json-tools:json-schema-validator:2.2.14",
-            "com.google.code.gson:gson:2.9.1",
+            "com.google.code.gson:gson:2.11.0",
             "com.google.guava:guava:32.0.1-jre",
             "com.google.protobuf:protobuf-java:3.23.4",
             "com.google.protobuf:protobuf-java-util:3.23.4",

python/ray/_private/accelerators/intel_gpu.py

@@ -7,8 +7,10 @@
 
 logger = logging.getLogger(__name__)
 
+ZE_AFFINITY_MASK_ENV_VAR = "ZE_AFFINITY_MASK"
+NOSET_ZE_AFFINITY_MASK_ENV_VAR = "RAY_EXPERIMENTAL_NOSET_ZE_AFFINITY_MASK"
+
 ONEAPI_DEVICE_SELECTOR_ENV_VAR = "ONEAPI_DEVICE_SELECTOR"
-NOSET_ONEAPI_DEVICE_SELECTOR_ENV_VAR = "RAY_EXPERIMENTAL_NOSET_ONEAPI_DEVICE_SELECTOR"
 ONEAPI_DEVICE_BACKEND_TYPE = "level_zero"
 ONEAPI_DEVICE_TYPE = "gpu"
 
@@ -22,25 +24,29 @@ def get_resource_name() -> str:
 
     @staticmethod
     def get_visible_accelerator_ids_env_var() -> str:
-        return ONEAPI_DEVICE_SELECTOR_ENV_VAR
+        return ZE_AFFINITY_MASK_ENV_VAR
 
     @staticmethod
     def get_current_process_visible_accelerator_ids() -> Optional[List[str]]:
-        oneapi_visible_devices = os.environ.get(
-            IntelGPUAcceleratorManager.get_visible_accelerator_ids_env_var(), None
-        )
+        # Primary: ZE_AFFINITY_MASK uses bare IDs ("0,1,2"), like CUDA_VISIBLE_DEVICES.
+        ze_mask = os.environ.get(ZE_AFFINITY_MASK_ENV_VAR, None)
+        if ze_mask is not None:
+            if ze_mask == "":
+                return []
+            return list(ze_mask.split(","))
+
+        # Fallback: ONEAPI_DEVICE_SELECTOR for backward compatibility.
+        oneapi_visible_devices = os.environ.get(ONEAPI_DEVICE_SELECTOR_ENV_VAR, None)
         if oneapi_visible_devices is None:
             return None
-
-        if oneapi_visible_devices == "":
-            return []
-
-        if oneapi_visible_devices == "NoDevFiles":
+        if oneapi_visible_devices == "" or oneapi_visible_devices == "NoDevFiles":
             return []
 
         prefix = ONEAPI_DEVICE_BACKEND_TYPE + ":"
-
-        return list(oneapi_visible_devices.split(prefix)[1].split(","))
+        if prefix in oneapi_visible_devices:
+            return list(oneapi_visible_devices.split(prefix)[1].split(","))
+        # bare IDs without prefix (e.g. "0,1") — accepted as-is
+        return list(oneapi_visible_devices.split(","))
 
     @staticmethod
     def get_current_node_num_accelerators() -> int:
@@ -95,10 +101,14 @@ def validate_resource_request_quantity(
     def set_current_process_visible_accelerator_ids(
         visible_xpu_devices: List[str],
     ) -> None:
-        if env_bool(NOSET_ONEAPI_DEVICE_SELECTOR_ENV_VAR, False):
+        if env_bool(NOSET_ZE_AFFINITY_MASK_ENV_VAR, False):
             return
 
-        prefix = ONEAPI_DEVICE_BACKEND_TYPE + ":"
-        os.environ[
-            IntelGPUAcceleratorManager.get_visible_accelerator_ids_env_var()
-        ] = prefix + ",".join([str(i) for i in visible_xpu_devices])
+        # ZE_AFFINITY_MASK masks devices at the Level Zero driver, below oneAPI/SYCL
+        # and torch-xpu, so it fully restricts visibility on its own. It uses bare
+        # IDs ("0,1,2") like CUDA_VISIBLE_DEVICES. Ray only sets this one env var:
+        # writing ONEAPI_DEVICE_SELECTOR too would collide with frameworks (e.g.
+        # SGLang) that manage it themselves, and would need a separate save/restore.
+        os.environ[ZE_AFFINITY_MASK_ENV_VAR] = ",".join(
+            [str(i) for i in visible_xpu_devices]
+        )