gpustack-runtime 0.2.1__tar.gz → 0.2.2__tar.gz

{gpustack_runtime-0.2.1 → gpustack_runtime-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpustack-runtime
-Version: 0.2.1
+Version: 0.2.2
 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-0.2.1 → gpustack_runtime-0.2.2}/gpustack_runtime/_version.py

@@ -27,8 +27,8 @@ version_tuple: VERSION_TUPLE
 __commit_id__: COMMIT_ID
 commit_id: COMMIT_ID
 
-__version__ = version = '0.2.1'
-__version_tuple__ = version_tuple = (0, 2, 1)
+__version__ = version = '0.2.2'
+__version_tuple__ = version_tuple = (0, 2, 2)
 try:
     from ._version_appendix import git_commit
     __commit_id__ = commit_id = git_commit

gpustack_runtime-0.2.2/gpustack_runtime/_version_appendix.py

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

{gpustack_runtime-0.2.1 → gpustack_runtime-0.2.2}/gpustack_runtime/deployer/__types__.py

@@ -1135,6 +1135,10 @@ class WorkloadStatus:
             Namespace for the workload.
         labels (dict[str, str] | None):
             Labels for the workload.
+        annotations (dict[str, str] | None):
+            Annotations for the workload.
+        state_message (str):
+            Message accompanying the current state.
         executable (list[WorkloadStatusOperation]):
             The operation for the executable containers of the workload.
         loggable (list[WorkloadStatusOperation]):
@@ -1161,6 +1165,15 @@ class WorkloadStatus:
     """
     Labels for the workload.
     """
+    annotations: dict[str, str] | None = field(default_factory=dict)
+    """
+    Annotations for the workload.
+    """
+    state_message: str = ""
+    """
+    Message accompanying the current state, e.g. the Kubernetes Pod's
+    status message on failure (such as a device-plugin admission rejection).
+    """
     executable: list[WorkloadStatusOperation] | None = field(default_factory=list)
     """
     The operation for the executable containers of the workload.

{gpustack_runtime-0.2.1 → gpustack_runtime-0.2.2}/gpustack_runtime/deployer/kuberentes.py

@@ -227,12 +227,18 @@ class KubernetesWorkloadStatus(WorkloadStatus):
             for k, v in (k_pod.metadata.labels or {}).items()
             if not k.startswith("runtime.gpustack.ai/")
         }
+        annotations = {
+            k: v
+            for k, v in (k_pod.metadata.annotations or {}).items()
+            if not k.startswith("runtime.gpustack.ai/")
+        }
 
         super().__init__(
             name=name,
             created_at=created_at,
             namespace=k_pod.metadata.namespace,
             labels=labels,
+            annotations=annotations,
             **kwargs,
         )
 
@@ -259,6 +265,11 @@ class KubernetesWorkloadStatus(WorkloadStatus):
             self.loggable.append(op)
 
         self.state = self.parse_state(k_pod)
+        if k_pod.status.message:
+            # Surface the Pod's status message, e.g. a device-plugin
+            # admission rejection ("UnexpectedAdmissionError: Allocate
+            # failed ...") on Failed Pods.
+            self.state_message = k_pod.status.message
 
 
 _NAME = "kubernetes"
@@ -267,6 +278,29 @@ Name of the Kubernetes deployer.
 """
 
 
+def _pin_pod_for_kueue(
+    pod: kubernetes.client.V1Pod,
+    node_name: str | None,
+) -> None:
+    """
+    Adjust node pinning for Kueue-gated Pods in place.
+
+    Kueue-gated Pods (queue label present) cannot set spec.nodeName:
+    the API forbids nodeName until all schedulingGates have been
+    cleared. Pin via nodeSelector instead; kube-scheduler places
+    the Pod on the same node once Kueue admits it.
+    """
+    if not node_name:
+        return
+    if "kueue.x-k8s.io/queue-name" not in (pod.metadata.labels or {}):
+        return
+    pod.spec.node_name = None
+    pod.spec.node_selector = {
+        **(pod.spec.node_selector or {}),
+        "kubernetes.io/hostname": node_name,
+    }
+
+
 class KubernetesDeployer(EndoscopicDeployer):
     """
     Deployer implementation for Kubernetes.
@@ -1191,6 +1225,8 @@ class KubernetesDeployer(EndoscopicDeployer):
 
         core_api = kubernetes.client.CoreV1Api(self._client)
         try:
+            _pin_pod_for_kueue(pod, self._node_name)
+
             pod = self._mutate_create_pod(pod)
             if envs.GPUSTACK_RUNTIME_DEPLOY_PRINT_CONVERSION:
                 clogger.info(

gpustack_runtime-0.2.2/tests/gpustack_runtime/deployer/test_workload_status.py

@@ -0,0 +1,117 @@
+from datetime import datetime, timezone
+
+import kubernetes.client
+
+from gpustack_runtime.deployer.__types__ import WorkloadStatus
+from gpustack_runtime.deployer.kuberentes import (
+    KubernetesWorkloadStatus,
+    _pin_pod_for_kueue,
+)
+
+
+def _pod(annotations=None) -> kubernetes.client.V1Pod:
+    return kubernetes.client.V1Pod(
+        metadata=kubernetes.client.V1ObjectMeta(
+            name="gpustack-test",
+            namespace="default",
+            creation_timestamp=datetime(2026, 7, 29, 8, 0, 0, tzinfo=timezone.utc),
+            labels={"app": "test"},
+            annotations=annotations,
+        ),
+        spec=kubernetes.client.V1PodSpec(
+            containers=[kubernetes.client.V1Container(name="run-0")],
+        ),
+        status=kubernetes.client.V1PodStatus(phase="Pending"),
+    )
+
+
+def test_kubernetes_workload_status_annotations():
+    status = KubernetesWorkloadStatus(
+        name="test",
+        k_pod=_pod(
+            annotations={
+                "device.gpustack.ai/accelerator.allocated": '{"run-0": {}}',
+                "runtime.gpustack.ai/component-run-0-name": "internal",
+            },
+        ),
+    )
+    # Third-party annotations (e.g. device-plugin allocation) are surfaced,
+    # runtime-internal annotations are filtered out.
+    assert status.annotations == {
+        "device.gpustack.ai/accelerator.allocated": '{"run-0": {}}',
+    }
+
+
+def test_kubernetes_workload_status_annotations_absent():
+    status = KubernetesWorkloadStatus(name="test", k_pod=_pod(annotations=None))
+    assert status.annotations == {}
+
+
+def test_workload_status_annotations_default_empty():
+    # Docker/Podman deployers don't populate annotations; the base default is empty.
+    status = WorkloadStatus(name="test", created_at="2026-07-29T08:00:00.000000Z")
+    assert status.annotations == {}
+
+
+def test_workload_status_annotations_json_roundtrip():
+    status = KubernetesWorkloadStatus(
+        name="test",
+        k_pod=_pod(annotations={"device.gpustack.ai/accelerator.allocated": "{}"}),
+    )
+    restored = WorkloadStatus.from_json(status.to_json())
+    assert restored.annotations == status.annotations
+
+
+def _pinning_pod(labels=None) -> kubernetes.client.V1Pod:
+    return kubernetes.client.V1Pod(
+        metadata=kubernetes.client.V1ObjectMeta(name="p", labels=labels or {}),
+        spec=kubernetes.client.V1PodSpec(
+            node_name="node-a",
+            containers=[kubernetes.client.V1Container(name="run-0")],
+        ),
+    )
+
+
+def test_pin_pod_for_kueue_switches_to_node_selector():
+    pod = _pinning_pod(labels={"kueue.x-k8s.io/queue-name": "q"})
+    _pin_pod_for_kueue(pod, "node-a")
+    assert pod.spec.node_name is None
+    assert pod.spec.node_selector == {"kubernetes.io/hostname": "node-a"}
+
+
+def test_pin_pod_for_kueue_merges_existing_node_selector():
+    pod = _pinning_pod(labels={"kueue.x-k8s.io/queue-name": "q"})
+    pod.spec.node_selector = {"disk": "ssd"}
+    _pin_pod_for_kueue(pod, "node-a")
+    assert pod.spec.node_selector == {
+        "disk": "ssd",
+        "kubernetes.io/hostname": "node-a",
+    }
+
+
+def test_pin_pod_for_kueue_noop_without_queue_label():
+    pod = _pinning_pod()
+    _pin_pod_for_kueue(pod, "node-a")
+    assert pod.spec.node_name == "node-a"
+    assert pod.spec.node_selector is None
+
+
+def test_pin_pod_for_kueue_noop_without_node_name():
+    pod = _pinning_pod(labels={"kueue.x-k8s.io/queue-name": "q"})
+    _pin_pod_for_kueue(pod, None)
+    assert pod.spec.node_name == "node-a"
+
+
+def test_kubernetes_workload_status_state_message():
+    pod = _pod()
+    pod.status = kubernetes.client.V1PodStatus(
+        phase="Failed",
+        message="Allocate failed due to no enough GPU devices",
+    )
+    status = KubernetesWorkloadStatus(name="test", k_pod=pod)
+    assert status.state_message == "Allocate failed due to no enough GPU devices"
+
+
+def test_kubernetes_workload_status_state_message_default():
+    status = KubernetesWorkloadStatus(name="test", k_pod=_pod())
+    assert status.state_message == ""

gpustack_runtime-0.2.1/gpustack_runtime/_version_appendix.py

@@ -1 +0,0 @@
-git_commit = "1d1e894"