gpustack-runtime 0.1.41.post1__py3-none-any.whl → 0.1.41.post3__py3-none-any.whl

gpustack_runtime/_version.py

@@ -27,8 +27,8 @@ version_tuple: VERSION_TUPLE
 __commit_id__: COMMIT_ID
 commit_id: COMMIT_ID
 
-__version__ = version = '0.1.41.post1'
-__version_tuple__ = version_tuple = (0, 1, 41, 'post1')
+__version__ = version = '0.1.41.post3'
+__version_tuple__ = version_tuple = (0, 1, 41, 'post3')
 try:
     from ._version_appendix import git_commit
     __commit_id__ = commit_id = git_commit

gpustack_runtime/_version_appendix.py

@@ -1 +1 @@
-git_commit = "8671a00"
+git_commit = "3840255"

gpustack_runtime/deployer/k8s/deviceplugin/__init__.py

@@ -65,7 +65,7 @@ async def serve_async(
         if not devices:
             continue
 
-        allocation_policy = _get_device_allocation_policy(manu)
+        allocation_policy = get_device_allocation_policy(manu)
         logger.info(
             "Using device allocation policy '%s' for manufacturer '%s'",
             allocation_policy,
@@ -110,6 +110,7 @@ async def serve_async(
                     device=dev,
                     id_by="index" if manu == ManufacturerEnum.ASCEND else "uuid",
                     allocation_policy=allocation_policy,
+                    max_allocations=envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_PER_DEVICE_MAX_ALLOCATIONS,
                 ),
             )
 
@@ -276,7 +277,23 @@ def is_kubelet_socket_accessible(
 
 
 @lru_cache
-def _get_device_allocation_policy(
+def get_resource_injection_policy() -> Literal["env", "kdp"]:
+    """
+    Get the resource injection policy (in lowercase) for the deployer.
+
+    Returns:
+        The resource injection policy.
+
+    """
+    policy = envs.GPUSTACK_RUNTIME_KUBERNETES_RESOURCE_INJECTION_POLICY.lower()
+    if policy != "auto":
+        return policy
+
+    return "kdp" if is_kubelet_socket_accessible() else "env"
+
+
+@lru_cache
+def get_device_allocation_policy(
     manufacturer: ManufacturerEnum,
 ) -> Literal["env", "cdi", "opaque"]:
     """
@@ -319,6 +336,8 @@ def _get_device_allocation_policy(
 
 __all__ = [
     "cdi_kind_to_kdp_resource",
+    "get_device_allocation_policy",
+    "get_resource_injection_policy",
     "is_kubelet_socket_accessible",
     "serve",
     "serve_async",

gpustack_runtime/deployer/k8s/deviceplugin/plugin.py

@@ -11,7 +11,6 @@ import grpc
 from grpc_interceptor import AsyncServerInterceptor
 from grpc_interceptor.exceptions import GrpcException
 
-from .... import envs
 from ....detector import Device, str_range_to_list
 from ...cdi import (
     generate_config,
@@ -135,7 +134,7 @@ class SharableDevicePlugin(PluginServer, DevicePluginServicer):
         device: Device,
         id_by: Literal["uuid", "index"] = "uuid",
         allocation_policy: Literal["env", "cdi", "opaque"] = "cdi",
-        max_allocations: int | None = None,
+        max_allocations: int | None = 10,
     ):
         """
         Initializes the SharableDevicePlugin.
@@ -150,18 +149,12 @@ class SharableDevicePlugin(PluginServer, DevicePluginServicer):
                 Controls the device allocation policy.
             max_allocations:
                 Controls the maximum allocations per underlying device.
-                If None, uses the environment variable `GPUSTACK_RUNTIME_KUBERNETES_KDP_PER_DEVICE_MAX_ALLOCATIONS`.
 
         """
         self._device = device
         self._id_by = id_by
         self._allocation_policy = allocation_policy
-        self._max_allocations = max_allocations
-        if not self._max_allocations:
-            self._max_allocations = (
-                envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_PER_DEVICE_MAX_ALLOCATIONS
-            )
-        self._max_allocations = max(self._max_allocations, 1)
+        self._max_allocations = max(max_allocations, 1)
         self._cdi_kind = manufacturer_to_cdi_kind(device.manufacturer)
         self._runtime_env = manufacturer_to_runtime_env(device.manufacturer)
         self._kdp_resource = cdi_kind_to_kdp_resource(
@@ -425,8 +418,7 @@ class SharableDevicePlugin(PluginServer, DevicePluginServicer):
         self,
         req: ContainerAllocateRequest,
     ) -> ContainerAllocateResponse:
-        policy = envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_DEVICE_ALLOCATION_POLICY.lower()
-
+        policy = self._allocation_policy
         request_dp_device_ids = req.devices_ids
 
         # CDI device allocation.