torchx-nightly 2025.8.5__py3-none-any.whl → 2026.1.11__py3-none-any.whl

torchx/schedulers/kubernetes_scheduler.py

@@ -27,10 +27,81 @@ Install Volcano:
 See the
 `Volcano Quickstart <https://github.com/volcano-sh/volcano>`_
 for more information.
+
+Pod Overlay
+===========
+
+You can overlay arbitrary Kubernetes Pod fields on generated pods by setting
+the ``kubernetes`` metadata on your role. The value can be:
+
+- A dict with the overlay structure
+- A resource URI pointing to a YAML file (e.g. ``file://``, ``s3://``, ``gs://``)
+
+Merge semantics:
+- **dict**: recursive merge (upsert)
+- **list**: append by default, replace if tuple (Python) or ``!!python/tuple`` tag (YAML)
+- **primitives**: replace
+
+.. code:: python
+
+    from torchx.specs import Role
+
+    # Dict overlay - lists append, tuples replace
+    role = Role(
+        name="trainer",
+        image="my-image:latest",
+        entrypoint="train.py",
+        metadata={
+            "kubernetes": {
+                "spec": {
+                    "nodeSelector": {"gpu": "true"},
+                    "tolerations": [{"key": "nvidia.com/gpu", "operator": "Exists"}],  # appends
+                    "volumes": ({"name": "my-volume", "emptyDir": {}},)  # replaces
+                }
+            }
+        }
+    )
+
+    # File URI overlay
+    role = Role(
+        name="trainer",
+        image="my-image:latest",
+        entrypoint="train.py",
+        metadata={
+            "kubernetes": "file:///path/to/pod_overlay.yaml"
+        }
+    )
+
+CLI usage with builtin components:
+
+.. code:: bash
+
+    $ torchx run --scheduler kubernetes dist.ddp \\
+        --metadata kubernetes=file:///path/to/pod_overlay.yaml \\
+        --script train.py
+
+Example ``pod_overlay.yaml``:
+
+.. code:: yaml
+
+    spec:
+      nodeSelector:
+        node.kubernetes.io/instance-type: p4d.24xlarge
+      tolerations:
+        - key: nvidia.com/gpu
+          operator: Exists
+          effect: NoSchedule
+      volumes: !!python/tuple
+        - name: my-volume
+          emptyDir: {}
+
+The overlay is deep-merged with the generated pod, preserving existing fields
+and adding or overriding specified ones.
 """
 
 import json
 import logging
+import re
 import warnings
 from dataclasses import dataclass
 from datetime import datetime
@@ -45,6 +116,7 @@ from typing import (
     Tuple,
     TYPE_CHECKING,
     TypedDict,
+    Union,
 )
 
 import torchx
@@ -77,7 +149,6 @@ from torchx.specs.api import (
 from torchx.util.strings import normalize_str
 from torchx.workspace.docker_workspace import DockerWorkspaceMixin
 
-
 if TYPE_CHECKING:
     from docker import DockerClient
     from kubernetes.client import ApiClient, CustomObjectsApi
@@ -87,6 +158,7 @@ if TYPE_CHECKING:
     )
     from kubernetes.client.rest import ApiException
 
+
 logger: logging.Logger = logging.getLogger(__name__)
 
 # Kubernetes reserves a small amount of resources per host for the system. For
@@ -97,6 +169,40 @@ logger: logging.Logger = logging.getLogger(__name__)
 RESERVED_MILLICPU = 100
 RESERVED_MEMMB = 1024
 
+
+def _apply_pod_overlay(pod: "V1Pod", overlay: Dict[str, Any]) -> None:
+    """Apply overlay dict to V1Pod object, merging nested fields.
+
+    Merge semantics:
+    - dict: upsert (recursive merge)
+    - list: append by default, replace if tuple
+    - primitives: replace
+    """
+    from kubernetes import client
+
+    api = client.ApiClient()
+    pod_dict = api.sanitize_for_serialization(pod)
+
+    def deep_merge(base: Dict[str, Any], overlay: Dict[str, Any]) -> None:
+        for key, value in overlay.items():
+            if isinstance(value, dict) and key in base and isinstance(base[key], dict):
+                deep_merge(base[key], value)
+            elif isinstance(value, tuple):
+                base[key] = list(value)
+            elif (
+                isinstance(value, list) and key in base and isinstance(base[key], list)
+            ):
+                base[key].extend(value)
+            else:
+                base[key] = value
+
+    deep_merge(pod_dict, overlay)
+
+    merged_pod = api._ApiClient__deserialize(pod_dict, "V1Pod")
+    pod.spec = merged_pod.spec
+    pod.metadata = merged_pod.metadata
+
+
 RETRY_POLICIES: Mapping[str, Iterable[Mapping[str, str]]] = {
     RetryPolicy.REPLICA: [],
     RetryPolicy.APPLICATION: [
@@ -188,7 +294,14 @@ def sanitize_for_serialization(obj: object) -> object:
     return api.sanitize_for_serialization(obj)
 
 
-def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod":
+def role_to_pod(
+    name: str,
+    role: Role,
+    service_account: Optional[str],
+    reserved_millicpu: int = RESERVED_MILLICPU,
+    reserved_memmb: int = RESERVED_MEMMB,
+    efa_device_count: Optional[int] = None,
+) -> "V1Pod":
     from kubernetes.client.models import (  # noqa: F811 redefinition of unused
         V1Container,
         V1ContainerPort,
@@ -218,18 +331,29 @@ def role_to_pod(name: str, role: Role, service_account: Optional[str]) -> "V1Pod
     if resource.cpu > 0:
         mcpu = int(resource.cpu * 1000)
         limits["cpu"] = f"{mcpu}m"
-        request_mcpu = max(mcpu - RESERVED_MILLICPU, 0)
+        request_mcpu = max(mcpu - reserved_millicpu, 0)
         requests["cpu"] = f"{request_mcpu}m"
     if resource.memMB > 0:
         limits["memory"] = f"{int(resource.memMB)}M"
-        request_memMB = max(int(resource.memMB) - RESERVED_MEMMB, 0)
+        request_memMB = max(int(resource.memMB) - reserved_memmb, 0)
         requests["memory"] = f"{request_memMB}M"
     if resource.gpu > 0:
         requests["nvidia.com/gpu"] = limits["nvidia.com/gpu"] = str(resource.gpu)
 
+    EFA_DEVICE = "vpc.amazonaws.com/efa"
     for device_name, device_limit in resource.devices.items():
         limits[device_name] = str(device_limit)
 
+    # Handle EFA device count override:
+    # - None (default): use whatever count is in the resource spec (already added above)
+    # - 0: remove EFA devices entirely
+    # - N > 0: set EFA device count to N (override or add)
+    if efa_device_count is not None:
+        if efa_device_count == 0:
+            limits.pop(EFA_DEVICE, None)
+        else:
+            limits[EFA_DEVICE] = str(efa_device_count)
+
     resources = V1ResourceRequirements(
         limits=limits,
         requests=requests,
@@ -369,7 +493,10 @@ def app_to_resource(
     queue: str,
     service_account: Optional[str],
     priority_class: Optional[str] = None,
-) -> Dict[str, object]:
+    reserved_millicpu: int = RESERVED_MILLICPU,
+    reserved_memmb: int = RESERVED_MEMMB,
+    efa_device_count: Optional[int] = None,
+) -> Dict[str, Any]:
     """
     app_to_resource creates a volcano job kubernetes resource definition from
     the provided AppDef. The resource definition can be used to launch the
@@ -399,8 +526,27 @@ def app_to_resource(
             replica_role = values.apply(role)
             if role_idx == 0 and replica_id == 0:
                 replica_role.env["TORCHX_RANK0_HOST"] = "localhost"
-
-            pod = role_to_pod(name, replica_role, service_account)
+            replica_role.env["TORCHX_IMAGE"] = replica_role.image
+
+            pod = role_to_pod(
+                name,
+                replica_role,
+                service_account,
+                reserved_millicpu,
+                reserved_memmb,
+                efa_device_count,
+            )
+            if k8s_metadata := role.metadata.get("kubernetes"):
+                if isinstance(k8s_metadata, str):
+                    import fsspec
+
+                    with fsspec.open(k8s_metadata, "r") as f:
+                        k8s_metadata = yaml.unsafe_load(f)
+                elif not isinstance(k8s_metadata, dict):
+                    raise ValueError(
+                        f"metadata['kubernetes'] must be a dict or resource URI, got {type(k8s_metadata)}"
+                    )
+                _apply_pod_overlay(pod, k8s_metadata)
             pod.metadata.labels.update(
                 pod_labels(
                     app=app,
@@ -443,7 +589,7 @@ does NOT support retries correctly. More info: https://github.com/volcano-sh/vol
     if priority_class is not None:
         job_spec["priorityClassName"] = priority_class
 
-    resource: Dict[str, object] = {
+    resource: Dict[str, Any] = {
         "apiVersion": "batch.volcano.sh/v1alpha1",
         "kind": "Job",
         "metadata": {"name": f"{unique_app_id}"},
@@ -455,7 +601,7 @@ does NOT support retries correctly. More info: https://github.com/volcano-sh/vol
 @dataclass
 class KubernetesJob:
     images_to_push: Dict[str, Tuple[str, str]]
-    resource: Dict[str, object]
+    resource: Dict[str, Any]
 
     def __str__(self) -> str:
         return yaml.dump(sanitize_for_serialization(self.resource))
@@ -470,12 +616,13 @@ class KubernetesOpts(TypedDict, total=False):
     image_repo: Optional[str]
     service_account: Optional[str]
     priority_class: Optional[str]
+    validate_spec: Optional[bool]
+    reserved_millicpu: Optional[int]
+    reserved_memmb: Optional[int]
+    efa_device_count: Optional[int]
 
 
-class KubernetesScheduler(
-    DockerWorkspaceMixin,
-    Scheduler[KubernetesOpts, AppDef, AppDryRunInfo[KubernetesJob]],
-):
+class KubernetesScheduler(DockerWorkspaceMixin, Scheduler[KubernetesOpts]):
     """
     KubernetesScheduler is a TorchX scheduling interface to Kubernetes.
 
@@ -485,7 +632,7 @@ class KubernetesScheduler(
     For installation instructions see: https://github.com/volcano-sh/volcano
 
     This has been confirmed to work with Volcano v1.3.0 and Kubernetes versions
-    v1.18-1.21. See https://github.com/pytorch/torchx/issues/120 which is
+    v1.18-1.21. See https://github.com/meta-pytorch/torchx/issues/120 which is
     tracking Volcano support for Kubernetes v1.22.
 
     .. note::
@@ -503,6 +650,16 @@ class KubernetesScheduler(
         $ torchx status kubernetes://torchx_user/1234
         ...
 
+    **Cancellation**
+
+    Canceling a job aborts it while preserving the job spec for inspection
+    and cloning via kubectl apply. Use the delete command to remove the job entirely:
+
+    .. code-block:: bash
+
+        $ torchx cancel kubernetes://namespace/jobname  # abort, preserves spec
+        $ torchx delete kubernetes://namespace/jobname  # delete completely
+
     **Config Options**
 
     .. runopts::
@@ -581,9 +738,14 @@ class KubernetesScheduler(
         if c is None:
             configuration = client.Configuration()
             try:
-                config.load_kube_config(client_configuration=configuration)
-            except config.ConfigException as e:
-                warnings.warn(f"failed to load kube config: {e}")
+                # Try in-cluster config first (for pods with ServiceAccount)
+                config.load_incluster_config(client_configuration=configuration)
+            except config.ConfigException:
+                # Fall back to kubeconfig (for local development)
+                try:
+                    config.load_kube_config(client_configuration=configuration)
+                except config.ConfigException as e:
+                    warnings.warn(f"failed to load kube config: {e}", stacklevel=2)
 
             c = self._client = client.ApiClient(configuration)
 
@@ -635,7 +797,7 @@ class KubernetesScheduler(
             else:
                 raise
 
-        return f'{namespace}:{resp["metadata"]["name"]}'
+        return f"{namespace}:{resp['metadata']['name']}"
 
     def _submit_dryrun(
         self, app: AppDef, cfg: KubernetesOpts
@@ -657,7 +819,56 @@ class KubernetesScheduler(
             priority_class, str
         ), "priority_class must be a str"
 
-        resource = app_to_resource(app, queue, service_account, priority_class)
+        reserved_millicpu = cfg.get("reserved_millicpu", RESERVED_MILLICPU)
+        assert isinstance(reserved_millicpu, int), "reserved_millicpu must be an int"
+
+        reserved_memmb = cfg.get("reserved_memmb", RESERVED_MEMMB)
+        assert isinstance(reserved_memmb, int), "reserved_memmb must be an int"
+
+        efa_device_count = cfg.get("efa_device_count")
+        assert efa_device_count is None or isinstance(
+            efa_device_count, int
+        ), "efa_device_count must be an int or None"
+
+        resource = app_to_resource(
+            app,
+            queue,
+            service_account,
+            priority_class,
+            reserved_millicpu,
+            reserved_memmb,
+            efa_device_count,
+        )
+
+        if cfg.get("validate_spec"):
+            try:
+                self._custom_objects_api().create_namespaced_custom_object(
+                    group="batch.volcano.sh",
+                    version="v1alpha1",
+                    namespace=cfg.get("namespace") or "default",
+                    plural="jobs",
+                    body=resource,
+                    dry_run="All",
+                )
+            except Exception as e:
+                from kubernetes.client.rest import ApiException
+
+                if isinstance(e, ApiException):
+                    raise ValueError(f"Invalid job spec: {e.reason}") from e
+                raise
+
+            job_name = resource["metadata"]["name"]
+            for task in resource["spec"]["tasks"]:
+                task_name = task["name"]
+                replicas = task.get("replicas", 1)
+                max_index = replicas - 1
+                pod_name = f"{job_name}-{task_name}-{max_index}"
+                if len(pod_name) > 63:
+                    raise ValueError(
+                        f"Pod name '{pod_name}' ({len(pod_name)} chars) exceeds 63 character limit. "
+                        f"Shorten app.name or role names"
+                    )
+
         req = KubernetesJob(
             resource=resource,
             images_to_push=images_to_push,
@@ -669,6 +880,31 @@ class KubernetesScheduler(
         pass
 
     def _cancel_existing(self, app_id: str) -> None:
+        """
+        Abort a Volcano job while preserving the spec for inspection.
+        """
+        namespace, name = app_id.split(":")
+        vcjob = self._custom_objects_api().get_namespaced_custom_object(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace=namespace,
+            plural="jobs",
+            name=name,
+        )
+        vcjob["status"]["state"]["phase"] = "Aborted"
+        self._custom_objects_api().replace_namespaced_custom_object_status(
+            group="batch.volcano.sh",
+            version="v1alpha1",
+            namespace=namespace,
+            plural="jobs",
+            name=name,
+            body=vcjob,
+        )
+
+    def _delete_existing(self, app_id: str) -> None:
+        """
+        Delete a Volcano job completely from the cluster.
+        """
         namespace, name = app_id.split(":")
         self._custom_objects_api().delete_namespaced_custom_object(
             group="batch.volcano.sh",
@@ -702,19 +938,52 @@ class KubernetesScheduler(
             type_=str,
             help="The name of the PriorityClass to set on the job specs",
         )
+        opts.add(
+            "validate_spec",
+            type_=bool,
+            help="Validate job spec using Kubernetes API dry-run before submission",
+            default=True,
+        )
+        opts.add(
+            "reserved_millicpu",
+            type_=int,
+            help="Amount of CPU in millicores to reserve for Kubernetes system overhead (default: 100)",
+            default=RESERVED_MILLICPU,
+        )
+        opts.add(
+            "reserved_memmb",
+            type_=int,
+            help="Amount of memory in MB to reserve for Kubernetes system overhead (default: 1024)",
+            default=RESERVED_MEMMB,
+        )
+        opts.add(
+            "efa_device_count",
+            type_=int,
+            help="EFA device count override: None/unset=use resource spec, "
+            "0=remove EFA, N>0=set EFA count to N",
+            default=None,
+        )
         return opts
 
     def describe(self, app_id: str) -> Optional[DescribeAppResponse]:
+        from kubernetes import client
+        from kubernetes.client.rest import ApiException
+
         namespace, name = app_id.split(":")
         roles = {}
         roles_statuses = {}
-        resp = self._custom_objects_api().get_namespaced_custom_object_status(
-            group="batch.volcano.sh",
-            version="v1alpha1",
-            namespace=namespace,
-            plural="jobs",
-            name=name,
-        )
+        try:
+            resp = self._custom_objects_api().get_namespaced_custom_object_status(
+                group="batch.volcano.sh",
+                version="v1alpha1",
+                namespace=namespace,
+                plural="jobs",
+                name=name,
+            )
+        except ApiException as e:
+            if e.status == 404:
+                return None
+            raise
         status = resp.get("status")
         if status:
             state_str = status["state"]["phase"]
@@ -723,18 +992,44 @@ class KubernetesScheduler(
             TASK_STATUS_COUNT = "taskStatusCount"
 
             if TASK_STATUS_COUNT in status:
-                for name, status in status[TASK_STATUS_COUNT].items():
-                    role, _, idx = name.rpartition("-")
+                for task_name, task_status in status[TASK_STATUS_COUNT].items():
+                    role, _, idx = task_name.rpartition("-")
 
-                    state_str = next(iter(status["phase"].keys()))
+                    state_str = next(iter(task_status["phase"].keys()))
                     state = TASK_STATE[state_str]
 
                     if role not in roles:
                         roles[role] = Role(name=role, num_replicas=0, image="")
                         roles_statuses[role] = RoleStatus(role, [])
                     roles[role].num_replicas += 1
+
+                    # Pod name follows the pattern: {job_name}-{task_name}-0
+                    # Get the pod to retrieve its IP address
+                    pod_name_k8s = f"{name}-{task_name}-0"
+                    hostname = ""
+                    try:
+                        core_api = client.CoreV1Api(self._api_client())
+                        pod = core_api.read_namespaced_pod(
+                            name=pod_name_k8s, namespace=namespace
+                        )
+                        pod_ip = pod.status.pod_ip
+
+                        if pod_ip is not None:
+                            # Convert IP to dashed format (e.g., 10.244.1.5 -> 10-244-1-5)
+                            pod_ip_dashed = pod_ip.replace(".", "-")
+
+                            # Kubernetes DNS = <pod-ip-dashed>.<namespace>.pod.cluster.local
+                            # Note: This will only be useful if the client using the IPs is in the cluster.
+                            hostname = f"{pod_ip_dashed}.{namespace}.pod.cluster.local"
+
+                    except ApiException:
+                        # Pod not found - hostname remains empty
+                        pass
+
                     roles_statuses[role].replicas.append(
-                        ReplicaStatus(id=int(idx), role=role, state=state, hostname="")
+                        ReplicaStatus(
+                            id=int(idx), role=role, state=state, hostname=hostname
+                        )
                     )
         else:
             app_state = AppState.UNKNOWN
@@ -778,7 +1073,10 @@ class KubernetesScheduler(
         core_api = client.CoreV1Api(self._api_client())
         if should_tail:
             w = watch.Watch()
-            iterator = w.stream(core_api.read_namespaced_pod_log, **args)
+            iterator = (
+                f"{line}\n"
+                for line in w.stream(core_api.read_namespaced_pod_log, **args)
+            )
         else:
             resp = core_api.read_namespaced_pod_log(**args)
             iterator = split_lines(resp)
@@ -823,13 +1121,34 @@ def create_scheduler(
 def pod_labels(
     app: AppDef, role_idx: int, role: Role, replica_id: int, app_id: str
 ) -> Dict[str, str]:
+
+    def clean(label_value: str) -> str:
+        # cleans the provided `label_value` to make it compliant
+        # to pod label specs as described in
+        # https://kubernetes.io/docs/concepts/overview/working-with-objects/labels/
+        #
+        # Valid label value:
+        # must be 63 characters or less (can be empty),
+        # unless empty, must begin and end with an alphanumeric character ([a-z0-9A-Z]),
+        # could contain dashes (-), underscores (_), dots (.), and alphanumerics between.
+
+        # Replace invalid characters (allow: alphanum, -, _, .) with "."
+        label_value = re.sub(r"[^A-Za-z0-9\-_.]", ".", label_value)
+        # Replace leading non-alphanumeric with "."
+        label_value = re.sub(r"^[^A-Za-z0-9]+", ".", label_value)
+        # Replace trailing non-alphanumeric with "."
+        label_value = re.sub(r"[^A-Za-z0-9]+$", ".", label_value)
+
+        # Trim to 63 characters
+        return label_value[:63]
+
     return {
-        LABEL_VERSION: torchx.__version__,
-        LABEL_APP_NAME: app.name,
+        LABEL_VERSION: clean(torchx.__version__),
+        LABEL_APP_NAME: clean(app.name),
         LABEL_ROLE_INDEX: str(role_idx),
-        LABEL_ROLE_NAME: role.name,
+        LABEL_ROLE_NAME: clean(role.name),
         LABEL_REPLICA_ID: str(replica_id),
-        LABEL_KUBE_APP_NAME: app.name,
+        LABEL_KUBE_APP_NAME: clean(app.name),
         LABEL_ORGANIZATION: "torchx.pytorch.org",
-        LABEL_UNIQUE_NAME: app_id,
+        LABEL_UNIQUE_NAME: clean(app_id),
     }

torchx/schedulers/local_scheduler.py

@@ -529,7 +529,7 @@ def _register_termination_signals() -> None:
         signal.signal(signal.SIGINT, _terminate_process_handler)
 
 
-class LocalScheduler(Scheduler[LocalOpts, AppDef, AppDryRunInfo[PopenRequest]]):
+class LocalScheduler(Scheduler[LocalOpts]):
     """
     Schedules on localhost. Containers are modeled as processes and
     certain properties of the container that are either not relevant
@@ -1159,6 +1159,7 @@ class LogIterator:
             self._check_finished()  # check to see if app has finished running
 
             if os.path.isfile(self._log_file):
+                time.sleep(0.1)  # fix timing issue
                 self._log_fp = open(
                     self._log_file,
                     mode="rt",

torchx/schedulers/lsf_scheduler.py

@@ -394,7 +394,7 @@ class LsfBsub:
 {self.materialize()}"""
 
 
-class LsfScheduler(Scheduler[LsfOpts, AppDef, AppDryRunInfo]):
+class LsfScheduler(Scheduler[LsfOpts]):
     """
     **Example: hello_world**