gpustack-runtime 0.2.2__tar.gz → 0.2.2.post1__tar.gz

{gpustack_runtime-0.2.2 → gpustack_runtime-0.2.2.post1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpustack-runtime
-Version: 0.2.2
+Version: 0.2.2.post1
 Summary: GPUStack Runtime is library for detecting GPU resources and launching GPU workloads.
 Project-URL: Homepage, https://github.com/gpustack/runtime
 Project-URL: Bug Tracker, https://github.com/gpustack/gpustack/issues

{gpustack_runtime-0.2.2 → gpustack_runtime-0.2.2.post1}/gpustack_runtime/_version.py

@@ -27,8 +27,8 @@ version_tuple: VERSION_TUPLE
 __commit_id__: COMMIT_ID
 commit_id: COMMIT_ID
 
-__version__ = version = '0.2.2'
-__version_tuple__ = version_tuple = (0, 2, 2)
+__version__ = version = '0.2.2.post1'
+__version_tuple__ = version_tuple = (0, 2, 2, 'post1')
 try:
     from ._version_appendix import git_commit
     __commit_id__ = commit_id = git_commit

gpustack_runtime-0.2.2.post1/gpustack_runtime/_version_appendix.py

@@ -0,0 +1 @@
+git_commit = "4d942f0"

{gpustack_runtime-0.2.2 → gpustack_runtime-0.2.2.post1}/gpustack_runtime/deployer/__types__.py

@@ -843,6 +843,8 @@ class WorkloadPlan(WorkloadSecurity):
             Name for the workload, it should be unique in the deployer.
         labels (dict[str, str] | None):
             Labels for the workload.
+        instance_type (str | None):
+            Name of the operator InstanceType the workload deploys onto.
         host_network (bool):
             Indicates if the containers of the workload use the host network.
         host_ipc (bool):
@@ -878,6 +880,13 @@ class WorkloadPlan(WorkloadSecurity):
     """
     Labels for the workload.
     """
+    instance_type: str | None = None
+    """
+    Name of the operator InstanceType the workload deploys onto.
+    When set (and Kueue admission is not disabled), the Kubernetes deployer
+    stamps the queue-name label of the InstanceType's entrance LocalQueue
+    onto the Pod so it is admitted by Kueue.
+    """
     host_network: bool = False
     """
     Indicates if the containers of the workload use the host network.

{gpustack_runtime-0.2.2 → gpustack_runtime-0.2.2.post1}/gpustack_runtime/deployer/kuberentes.py

@@ -278,6 +278,30 @@ Name of the Kubernetes deployer.
 """
 
 
+def _apply_instance_type_admission(
+    pod: kubernetes.client.V1Pod,
+    instance_type: str | None,
+) -> None:
+    """
+    Stamp the Kueue queue-name label of the InstanceType's entrance
+    LocalQueue onto the Pod, so the workload is admitted through the
+    operator's queue for that InstanceType. The LocalQueue name is the
+    FNV-1a 64 hash of the InstanceType name, mirroring the operator's
+    FormatLocalQueueName ("gpustack-fnv64-<fnv64a-hex>").
+
+    Skipped when no InstanceType is set, or when Kueue admission is
+    disabled via GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION —
+    then the Pod is scheduled by the plain Kubernetes scheduler.
+    """
+    if not instance_type:
+        return
+    if envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION:
+        return
+    pod.metadata.labels["kueue.x-k8s.io/queue-name"] = (
+        f"gpustack-fnv64-{fnv1a_64_hex(instance_type)}"
+    )
+
+
 def _pin_pod_for_kueue(
     pod: kubernetes.client.V1Pod,
     node_name: str | None,
@@ -1091,11 +1115,6 @@ class KubernetesDeployer(EndoscopicDeployer):
                             if kdp:
                                 # Request quantity of devices as resources for KDP to schedule on the correct nodes.
                                 resources.update(r_vs)
-                                # Request quantity of devices with Kueue admission.
-                                if not envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION:
-                                    pod.metadata.labels["kueue.x-k8s.io/queue-name"] = (
-                                        self._find_kueue_queue_name(ren)
-                                    )
                                 continue
                             # Request device via visible devices env.
                             container.env.append(
@@ -1119,11 +1138,6 @@ class KubernetesDeployer(EndoscopicDeployer):
                                 ] = ",".join(resource_values)
                                 # Request quantity of devices as resources for KDP to schedule on the correct nodes.
                                 resources.update(r_vs)
-                                # Request quantity of devices with Kueue admission.
-                                if not envs.GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION:
-                                    pod.metadata.labels["kueue.x-k8s.io/queue-name"] = (
-                                        self._find_kueue_queue_name(ren)
-                                    )
                                 continue
 
                             # Request all devices if privileged,
@@ -1225,6 +1239,7 @@ class KubernetesDeployer(EndoscopicDeployer):
 
         core_api = kubernetes.client.CoreV1Api(self._client)
         try:
+            _apply_instance_type_admission(pod, workload.instance_type)
             _pin_pod_for_kueue(pod, self._node_name)
 
             pod = self._mutate_create_pod(pod)
@@ -1583,31 +1598,6 @@ class KubernetesDeployer(EndoscopicDeployer):
             namespace=self_pod_namespace,
         )
 
-    def _find_kueue_queue_name(self, runtime_env: str) -> str:
-        manu = self.get_manufacturer(runtime_env)
-        core_api = kubernetes.client.CoreV1Api(self._client)
-        try:
-            # Iterate the labels of node to find the key as below:
-            # "feature.gpustack.ai/${manu}-${device-group-id}.profile-queue: <profile>",
-            # then combine the manufacturer, device group id and profile to get the queue name for Kueue admission.
-            #
-            # For example, for `feature.gpustack.ai/nvidia-tesla-t4.profile-queue: 12c-46g-1d`,
-            # the node key is `nvidia-tesla-t4`,
-            # the queue name is `gpustack-nvidia-tesla-t4-12c-46g-1d`.
-            node = core_api.read_node(name=self._node_name)
-            prefix = "feature.gpustack.ai/"
-            suffix = ".profile-queue"
-            for k, v in node.metadata.labels.items():
-                if k.startswith(f"{prefix}{manu}-") and k.endswith(suffix):
-                    node_key = k[len(prefix) : -len(suffix)]
-                    return f"gpustack-{node_key}-{v}"
-        except kubernetes.client.exceptions.ApiException as e:
-            msg = f"Failed to get Kueue queue name on node {self._node_name} for runtime environment {runtime_env}{_detail_api_call_error(e)}"
-            raise OperationError(msg) from e
-
-        msg = f"Failed to find Kueue queue name on node {self._node_name} for runtime environment {runtime_env}"
-        raise OperationError(msg)
-
     @_supported
     def _create(self, workload: WorkloadPlan):
         """

{gpustack_runtime-0.2.2 → gpustack_runtime-0.2.2.post1}/tests/gpustack_runtime/deployer/test_workload_status.py

@@ -2,6 +2,7 @@ from datetime import datetime, timezone
 
 import kubernetes.client
 
+from gpustack_runtime import envs
 from gpustack_runtime.deployer.__types__ import WorkloadStatus
 from gpustack_runtime.deployer.kuberentes import (
     KubernetesWorkloadStatus,
@@ -115,3 +116,40 @@ def test_kubernetes_workload_status_state_message():
 def test_kubernetes_workload_status_state_message_default():
     status = KubernetesWorkloadStatus(name="test", k_pod=_pod())
     assert status.state_message == ""
+
+
+from gpustack_runtime.deployer.kuberentes import (  # noqa: E402
+    _apply_instance_type_admission,
+)
+
+
+def test_apply_instance_type_admission_stamps_entrance_queue():
+    pod = _pinning_pod()
+    _apply_instance_type_admission(
+        pod,
+        "gpustack--nvidia-geforce-rtx-4090-linux-amd64",
+    )
+    # Vector observed on a live operator cluster.
+    assert (
+        pod.metadata.labels["kueue.x-k8s.io/queue-name"]
+        == "gpustack-fnv64-1a7145d0a5248d55"
+    )
+
+
+def test_apply_instance_type_admission_skipped_without_instance_type():
+    pod = _pinning_pod()
+    _apply_instance_type_admission(pod, None)
+    assert "kueue.x-k8s.io/queue-name" not in pod.metadata.labels
+
+
+def test_apply_instance_type_admission_disabled_by_env(monkeypatch):
+    monkeypatch.setenv("GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION", "true")
+
+    monkeypatch.setattr(
+        envs,
+        "GPUSTACK_RUNTIME_KUBERNETES_KDP_NO_KUEUE_ADMISSION",
+        True,
+    )
+    pod = _pinning_pod()
+    _apply_instance_type_admission(pod, "some-type")
+    assert "kueue.x-k8s.io/queue-name" not in pod.metadata.labels

gpustack_runtime-0.2.2/gpustack_runtime/_version_appendix.py

@@ -1 +0,0 @@
-git_commit = "3ea35d7"