torchx-nightly 2025.9.28__py3-none-any.whl → 2025.11.17__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
@@ -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: [
@@ -369,7 +475,7 @@ def app_to_resource(
     queue: str,
     service_account: Optional[str],
     priority_class: Optional[str] = None,
-) -> Dict[str, object]:
+) -> 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 +505,20 @@ def app_to_resource(
             replica_role = values.apply(role)
             if role_idx == 0 and replica_id == 0:
                 replica_role.env["TORCHX_RANK0_HOST"] = "localhost"
+            replica_role.env["TORCHX_IMAGE"] = replica_role.image
 
             pod = role_to_pod(name, replica_role, service_account)
+            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 +561,7 @@ 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 +573,7 @@ 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))
@@ -470,6 +588,7 @@ class KubernetesOpts(TypedDict, total=False):
     image_repo: Optional[str]
     service_account: Optional[str]
     priority_class: Optional[str]
+    validate_spec: Optional[bool]
 
 
 class KubernetesScheduler(
@@ -485,7 +604,7 @@ 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::
@@ -635,7 +754,7 @@ class KubernetesScheduler(
             else:
                 raise
 
-        return f'{namespace}:{resp["metadata"]["name"]}'
+        return f"{namespace}:{resp['metadata']['name']}"
 
     def _submit_dryrun(
         self, app: AppDef, cfg: KubernetesOpts
@@ -658,6 +777,36 @@ class KubernetesScheduler(
         ), "priority_class must be a str"
 
         resource = app_to_resource(app, queue, service_account, priority_class)
+
+        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,
@@ -702,19 +851,32 @@ 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,
+        )
         return opts
 
     def describe(self, app_id: str) -> Optional[DescribeAppResponse]:
+        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"]
@@ -823,13 +985,34 @@ 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/slurm_scheduler.py

@@ -73,6 +73,15 @@ def appstate_from_slurm_state(slurm_state: str) -> AppState:
     return SLURM_STATES.get(slurm_state, AppState.UNKNOWN)
 
 
+def get_appstate_from_job(job: dict[str, object]) -> AppState:
+    # Prior to slurm-23.11, job_state was a string and not a list
+    job_state = job.get("job_state", None)
+    if isinstance(job_state, list):
+        return appstate_from_slurm_state(job_state[0])
+    else:
+        return appstate_from_slurm_state(str(job_state))
+
+
 def version() -> Tuple[int, int]:
     """
     Uses ``sinfo --version`` to get the slurm version. If the command fails, it
@@ -666,7 +675,7 @@ class SlurmScheduler(
 
             entrypoint = job["command"]
             image = job["current_working_directory"]
-            state = appstate_from_slurm_state(job["job_state"][0])
+            state = get_appstate_from_job(job)
 
             job_resources = job["job_resources"]
 
@@ -881,7 +890,7 @@ class SlurmScheduler(
             out.append(
                 ListAppResponse(
                     app_id=str(job["job_id"]),
-                    state=SLURM_STATES[job["job_state"][0]],
+                    state=get_appstate_from_job(job),
                     name=job["name"],
                 )
             )

torchx/specs/__init__.py

@@ -12,7 +12,9 @@ used by components to define the apps which can then be launched via a TorchX
 scheduler or pipeline adapter.
 """
 import difflib
-from typing import Callable, Dict, Mapping, Optional
+
+import os
+from typing import Callable, Dict, Iterator, Mapping, Optional
 
 from torchx.specs.api import (
     ALL,
@@ -41,9 +43,11 @@ from torchx.specs.api import (
     RoleStatus,
     runopt,
     runopts,
+    TORCHX_HOME,
     UnknownAppException,
     UnknownSchedulerException,
     VolumeMount,
+    Workspace,
 )
 from torchx.specs.builders import make_app_handle, materialize_appdef, parse_mounts
 
@@ -53,6 +57,7 @@ from torchx.util.modules import import_attr
 
 GiB: int = 1024
 
+
 ResourceFactory = Callable[[], Resource]
 
 AWS_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
@@ -61,8 +66,10 @@ AWS_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
 GENERIC_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
     "torchx.specs.named_resources_generic", "NAMED_RESOURCES", default={}
 )
-FB_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
-    "torchx.specs.fb.named_resources", "NAMED_RESOURCES", default={}
+CUSTOM_NAMED_RESOURCES: Mapping[str, ResourceFactory] = import_attr(
+    os.environ.get("TORCHX_CUSTOM_NAMED_RESOURCES", "torchx.specs.fb.named_resources"),
+    "NAMED_RESOURCES",
+    default={},
 )
 
 
@@ -73,7 +80,7 @@ def _load_named_resources() -> Dict[str, Callable[[], Resource]]:
     for name, resource in {
         **GENERIC_NAMED_RESOURCES,
         **AWS_NAMED_RESOURCES,
-        **FB_NAMED_RESOURCES,
+        **CUSTOM_NAMED_RESOURCES,
         **resource_methods,
     }.items():
         materialized_resources[name] = resource
@@ -106,8 +113,22 @@ class _NamedResourcesLibrary:
     def __contains__(self, key: str) -> bool:
         return key in _named_resource_factories
 
-    def __iter__(self) -> None:
-        raise NotImplementedError("named resources doesn't support iterating")
+    def __iter__(self) -> Iterator[str]:
+        """Iterates through the names of the registered named_resources.
+
+        Usage:
+
+        .. doctest::
+
+            from torchx import specs
+
+            for resource_name in specs.named_resources:
+                resource = specs.resource(h=resource_name)
+                assert isinstance(resource, specs.Resource)
+
+        """
+        for key in _named_resource_factories:
+            yield (key)
 
 
 named_resources: _NamedResourcesLibrary = _NamedResourcesLibrary()
@@ -127,7 +148,7 @@ def resource(
 
     If ``h`` is specified then it is used to look up the
     resource specs from the list of registered named resources.
-    See `registering named resource <https://pytorch.org/torchx/latest/advanced.html#registering-named-resources>`_.
+    See `registering named resource <https://meta-pytorch.org/torchx/latest/advanced.html#registering-named-resources>`_.
 
     Otherwise a ``Resource`` object is created from the raw resource specs.
 
@@ -234,4 +255,6 @@ __all__ = [
     "torchx_run_args_from_json",
     "TorchXRunArgs",
     "ALL",
+    "TORCHX_HOME",
+    "Workspace",
 ]