skypilot-nightly 1.0.0.dev20250106__tar.gz → 1.0.0.dev20250108__tar.gz

{skypilot_nightly-1.0.0.dev20250106/skypilot_nightly.egg-info → skypilot_nightly-1.0.0.dev20250108}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: skypilot-nightly
-Version: 1.0.0.dev20250106
+Version: 1.0.0.dev20250108
 Summary: SkyPilot: An intercloud broker for the clouds
 Author: SkyPilot Team
 License: Apache 2.0
@@ -90,7 +90,7 @@ Requires-Dist: grpcio!=1.48.0,<=1.51.3,>=1.42.0; (python_version >= "3.10" and s
 Requires-Dist: protobuf!=3.19.5,>=3.15.3; extra == "remote"
 Requires-Dist: pydantic!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.*,!=2.4.*,<3; extra == "remote"
 Provides-Extra: runpod
-Requires-Dist: runpod>=1.5.1; extra == "runpod"
+Requires-Dist: runpod>=1.6.1; extra == "runpod"
 Provides-Extra: fluidstack
 Provides-Extra: cudo
 Requires-Dist: cudo-compute>=0.1.10; extra == "cudo"
@@ -140,7 +140,7 @@ Requires-Dist: grpcio!=1.48.0,<=1.51.3,>=1.32.0; (python_version < "3.10" and sy
 Requires-Dist: grpcio!=1.48.0,<=1.51.3,>=1.42.0; (python_version >= "3.10" and sys_platform != "darwin") and extra == "all"
 Requires-Dist: protobuf!=3.19.5,>=3.15.3; extra == "all"
 Requires-Dist: pydantic!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.*,!=2.4.*,<3; extra == "all"
-Requires-Dist: runpod>=1.5.1; extra == "all"
+Requires-Dist: runpod>=1.6.1; extra == "all"
 Requires-Dist: cudo-compute>=0.1.10; extra == "all"
 Requires-Dist: pydo>=0.3.0; extra == "all"
 Requires-Dist: azure-core>=1.24.0; extra == "all"

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/__init__.py

@@ -5,7 +5,7 @@ from typing import Optional
 import urllib.request
 
 # Replaced with the current commit when building the wheels.
-_SKYPILOT_COMMIT_SHA = '38a822ac6b553df0e784e559715ee4269c21f780'
+_SKYPILOT_COMMIT_SHA = '2fa37ec2a68bdd1e69f3726ca46aaeebcaf07a2d'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20250106'
+__version__ = '1.0.0.dev20250108'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/backends/cloud_vm_ray_backend.py

@@ -4216,11 +4216,20 @@ class CloudVmRayBackend(backends.Backend['CloudVmRayResourceHandle']):
         attempts = 0
         while True:
             logger.debug(f'instance statuses attempt {attempts + 1}')
-            node_status_dict = provision_lib.query_instances(
-                repr(cloud),
-                cluster_name_on_cloud,
-                config['provider'],
-                non_terminated_only=False)
+            try:
+                node_status_dict = provision_lib.query_instances(
+                    repr(cloud),
+                    cluster_name_on_cloud,
+                    config['provider'],
+                    non_terminated_only=False)
+            except Exception as e:  # pylint: disable=broad-except
+                if purge:
+                    logger.warning(
+                        f'Failed to query instances. Skipping since purge is '
+                        f'set. Details: '
+                        f'{common_utils.format_exception(e, use_bracket=True)}')
+                    break
+                raise
 
             unexpected_node_state: Optional[Tuple[str, str]] = None
             for node_id, node_status in node_status_dict.items():
@@ -4239,8 +4248,13 @@ class CloudVmRayBackend(backends.Backend['CloudVmRayResourceHandle']):
                 time.sleep(_TEARDOWN_WAIT_BETWEEN_ATTEMPS_SECONDS)
             else:
                 (node_id, node_status) = unexpected_node_state
-                raise RuntimeError(f'Instance {node_id} in unexpected state '
-                                   f'{node_status}.')
+                if purge:
+                    logger.warning(f'Instance {node_id} in unexpected '
+                                   f'state {node_status}. Skipping since purge '
+                                   'is set.')
+                    break
+                raise RuntimeError(f'Instance {node_id} in unexpected '
+                                   f'state {node_status}.')
 
         global_user_state.remove_cluster(handle.cluster_name,
                                          terminate=terminate)

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/clouds/service_catalog/data_fetchers/fetch_gcp.py

@@ -47,10 +47,6 @@ TPU_RETRY_CNT = 3
 TPU_V4_ZONES = ['us-central2-b']
 # TPU v3 pods are available in us-east1-d, but hidden in the skus.
 # We assume the TPU prices are the same as us-central1.
-# TPU v6e's pricing info is not available on the SKUs. However, in
-# https://cloud.google.com/tpu/pricing, it listed the price for 4 regions:
-# us-east1, us-east5, europe-west4, and asia-northeast1. We hardcode them here
-# and filtered out the other regions (us-central{1,2}, us-south1).
 HIDDEN_TPU_DF = pd.read_csv(
     io.StringIO(
         textwrap.dedent("""\
@@ -62,49 +58,10 @@ HIDDEN_TPU_DF = pd.read_csv(
  ,tpu-v3-512,1,,,tpu-v3-512,512.0,153.6,us-east1,us-east1-d
  ,tpu-v3-1024,1,,,tpu-v3-1024,1024.0,307.2,us-east1,us-east1-d
  ,tpu-v3-2048,1,,,tpu-v3-2048,2048.0,614.4,us-east1,us-east1-d
- ,tpu-v6e-1,1,,,tpu-v6e-1,2.7,,us-east5,us-east5-b
- ,tpu-v6e-1,1,,,tpu-v6e-1,2.7,,us-east5,us-east5-c
- ,tpu-v6e-1,1,,,tpu-v6e-1,2.97,,europe-west4,europe-west4-a
- ,tpu-v6e-1,1,,,tpu-v6e-1,3.24,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-1,1,,,tpu-v6e-1,2.7,,us-east1,us-east1-d
- ,tpu-v6e-4,1,,,tpu-v6e-4,10.8,,us-east5,us-east5-b
- ,tpu-v6e-4,1,,,tpu-v6e-4,10.8,,us-east5,us-east5-c
- ,tpu-v6e-4,1,,,tpu-v6e-4,11.88,,europe-west4,europe-west4-a
- ,tpu-v6e-4,1,,,tpu-v6e-4,12.96,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-4,1,,,tpu-v6e-4,10.8,,us-east1,us-east1-d
- ,tpu-v6e-8,1,,,tpu-v6e-8,21.6,,us-east5,us-east5-b
- ,tpu-v6e-8,1,,,tpu-v6e-8,21.6,,us-east5,us-east5-c
- ,tpu-v6e-8,1,,,tpu-v6e-8,23.76,,europe-west4,europe-west4-a
- ,tpu-v6e-8,1,,,tpu-v6e-8,25.92,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-8,1,,,tpu-v6e-8,21.6,,us-east1,us-east1-d
- ,tpu-v6e-16,1,,,tpu-v6e-16,43.2,,us-east5,us-east5-b
- ,tpu-v6e-16,1,,,tpu-v6e-16,43.2,,us-east5,us-east5-c
- ,tpu-v6e-16,1,,,tpu-v6e-16,47.52,,europe-west4,europe-west4-a
- ,tpu-v6e-16,1,,,tpu-v6e-16,51.84,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-16,1,,,tpu-v6e-16,43.2,,us-east1,us-east1-d
- ,tpu-v6e-32,1,,,tpu-v6e-32,86.4,,us-east5,us-east5-b
- ,tpu-v6e-32,1,,,tpu-v6e-32,86.4,,us-east5,us-east5-c
- ,tpu-v6e-32,1,,,tpu-v6e-32,95.04,,europe-west4,europe-west4-a
- ,tpu-v6e-32,1,,,tpu-v6e-32,103.68,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-32,1,,,tpu-v6e-32,86.4,,us-east1,us-east1-d
- ,tpu-v6e-64,1,,,tpu-v6e-64,172.8,,us-east5,us-east5-b
- ,tpu-v6e-64,1,,,tpu-v6e-64,172.8,,us-east5,us-east5-c
- ,tpu-v6e-64,1,,,tpu-v6e-64,190.08,,europe-west4,europe-west4-a
- ,tpu-v6e-64,1,,,tpu-v6e-64,207.36,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-64,1,,,tpu-v6e-64,172.8,,us-east1,us-east1-d
- ,tpu-v6e-128,1,,,tpu-v6e-128,345.6,,us-east5,us-east5-b
- ,tpu-v6e-128,1,,,tpu-v6e-128,345.6,,us-east5,us-east5-c
- ,tpu-v6e-128,1,,,tpu-v6e-128,380.16,,europe-west4,europe-west4-a
- ,tpu-v6e-128,1,,,tpu-v6e-128,414.72,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-128,1,,,tpu-v6e-128,345.6,,us-east1,us-east1-d
- ,tpu-v6e-256,1,,,tpu-v6e-256,691.2,,us-east5,us-east5-b
- ,tpu-v6e-256,1,,,tpu-v6e-256,691.2,,us-east5,us-east5-c
- ,tpu-v6e-256,1,,,tpu-v6e-256,760.32,,europe-west4,europe-west4-a
- ,tpu-v6e-256,1,,,tpu-v6e-256,829.44,,asia-northeast1,asia-northeast1-b
- ,tpu-v6e-256,1,,,tpu-v6e-256,691.2,,us-east1,us-east1-d
  """)))
 
-TPU_V6E_MISSING_REGIONS = ['us-central1', 'us-central2', 'us-south1']
+# TPU V6e price for us-central2 is missing in the SKUs.
+TPU_V6E_MISSING_REGIONS = ['us-central2']
 
 # TPU V5 is not visible in specific zones. We hardcode the missing zones here.
 # NOTE(dev): Keep the zones and the df in sync.
@@ -670,6 +627,8 @@ def get_tpu_df(gce_skus: List[Dict[str, Any]],
                 return 'TpuV5p'
             assert tpu_version == 'v5litepod', tpu_version
             return 'TpuV5e'
+        if tpu_version.startswith('v6e'):
+            return 'TpuV6e'
         return f'Tpu-{tpu_version}'
 
     def get_tpu_price(row: pd.Series, spot: bool) -> Optional[float]:
@@ -684,10 +643,10 @@ def get_tpu_df(gce_skus: List[Dict[str, Any]],
         # whether the TPU is a single device or a pod.
         # For TPU-v4, the pricing is uniform, and thus the pricing API
         # only provides the price of TPU-v4 pods.
-        # The price shown for v5 TPU is per chip hour, so there is no 'Pod'
-        # keyword in the description.
+        # The price shown for v5 & v6e TPU is per chip hour, so there is
+        # no 'Pod' keyword in the description.
         is_pod = ((num_cores > 8 or tpu_version == 'v4') and
-                  not tpu_version.startswith('v5'))
+                  not tpu_version.startswith('v5') and tpu_version != 'v6e')
 
         for sku in gce_skus + tpu_skus:
             if tpu_region not in sku['serviceRegions']:
@@ -718,7 +677,9 @@ def get_tpu_df(gce_skus: List[Dict[str, Any]],
             # for v5e. Reference here:
             # https://cloud.google.com/tpu/docs/v5p#using-accelerator-type
             # https://cloud.google.com/tpu/docs/v5e#tpu-v5e-config
-            core_per_sku = (1 if tpu_version == 'v5litepod' else
+            # v6e is also per chip price. Reference here:
+            # https://cloud.google.com/tpu/docs/v6e#configurations
+            core_per_sku = (1 if tpu_version in ['v5litepod', 'v6e'] else
                             2 if tpu_version == 'v5p' else 8)
             tpu_core_price = tpu_device_price / core_per_sku
             tpu_price = num_cores * tpu_core_price
@@ -738,8 +699,6 @@ def get_tpu_df(gce_skus: List[Dict[str, Any]],
             spot_str = 'spot ' if spot else ''
             print(f'The {spot_str}price of {tpu_name} in {tpu_region} is '
                   'not found in SKUs or hidden TPU price DF.')
-        # TODO(tian): Hack. Should investigate how to retrieve the price
-        # for TPU-v6e.
         if (tpu_name.startswith('tpu-v6e') and
                 tpu_region in TPU_V6E_MISSING_REGIONS):
             if not spot:

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/clouds/utils/oci_utils.py

@@ -10,6 +10,8 @@ History:
    from ubuntu 20.04 to ubuntu 22.04, including:
    - GPU: skypilot:gpu-ubuntu-2004 -> skypilot:gpu-ubuntu-2204
    - CPU: skypilot:cpu-ubuntu-2004 -> skypilot:cpu-ubuntu-2204
+ - Hysun He (hysun.he@oracle.com) @ Jan.01, 2025: Support reuse existing
+   VCN for SkyServe.
 """
 import os
 
@@ -109,8 +111,15 @@ class OCIConfig:
             ('oci', region, 'compartment_ocid'), default_compartment_ocid)
         return compartment
 
+    @classmethod
+    def get_vcn_ocid(cls, region):
+        # Will reuse the regional VCN if specified.
+        vcn = skypilot_config.get_nested(('oci', region, 'vcn_ocid'), None)
+        return vcn
+
     @classmethod
     def get_vcn_subnet(cls, region):
+        # Will reuse the subnet if specified.
         vcn = skypilot_config.get_nested(('oci', region, 'vcn_subnet'), None)
         return vcn
 

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/provision/docker_utils.py

@@ -38,6 +38,13 @@ class DockerLoginConfig:
     password: str
     server: str
 
+    def format_image(self, image: str) -> str:
+        """Format the image name with the server prefix."""
+        server_prefix = f'{self.server}/'
+        if not image.startswith(server_prefix):
+            return f'{server_prefix}{image}'
+        return image
+
     @classmethod
     def from_env_vars(cls, d: Dict[str, str]) -> 'DockerLoginConfig':
         return cls(
@@ -220,9 +227,7 @@ class DockerInitializer:
                 wait_for_docker_daemon=True)
             # We automatically add the server prefix to the image name if
             # the user did not add it.
-            server_prefix = f'{docker_login_config.server}/'
-            if not specific_image.startswith(server_prefix):
-                specific_image = f'{server_prefix}{specific_image}'
+            specific_image = docker_login_config.format_image(specific_image)
 
         if self.docker_config.get('pull_before_run', True):
             assert specific_image, ('Image must be included in config if ' +

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/provision/oci/query_utils.py

@@ -7,6 +7,8 @@ History:
    find_compartment: allow search subtree when find a compartment.
  - Hysun He (hysun.he@oracle.com) @ Nov.12, 2024: Add methods to
    Add/remove security rules: create_nsg_rules & remove_nsg
+ - Hysun He (hysun.he@oracle.com) @ Jan.01, 2025: Support reuse existing
+   VCN for SkyServe.
 """
 from datetime import datetime
 import functools
@@ -17,7 +19,6 @@ import traceback
 import typing
 from typing import List, Optional, Tuple
 
-from sky import exceptions
 from sky import sky_logging
 from sky.adaptors import common as adaptors_common
 from sky.adaptors import oci as oci_adaptor
@@ -496,26 +497,25 @@ class QueryHelper:
 
         compartment = cls.find_compartment(region)
 
-        list_vcns_resp = net_client.list_vcns(
-            compartment_id=compartment,
-            display_name=oci_utils.oci_config.VCN_NAME,
-            lifecycle_state='AVAILABLE',
-        )
-
-        if not list_vcns_resp:
-            raise exceptions.ResourcesUnavailableError(
-                'The VCN is not available')
+        vcn_id = oci_utils.oci_config.get_vcn_ocid(region)
+        if vcn_id is None:
+            list_vcns_resp = net_client.list_vcns(
+                compartment_id=compartment,
+                display_name=oci_utils.oci_config.VCN_NAME,
+                lifecycle_state='AVAILABLE',
+            )
 
-        # Get the primary vnic. The vnic might be an empty list for the
-        # corner case when the cluster was exited during provision.
-        if not list_vcns_resp.data:
-            return None
+            # Get the primary vnic. The vnic might be an empty list for the
+            # corner case when the cluster was exited during provision.
+            if not list_vcns_resp.data:
+                return None
 
-        vcn = list_vcns_resp.data[0]
+            vcn = list_vcns_resp.data[0]
+            vcn_id = vcn.id
 
         list_nsg_resp = net_client.list_network_security_groups(
             compartment_id=compartment,
-            vcn_id=vcn.id,
+            vcn_id=vcn_id,
             limit=1,
             display_name=nsg_name,
         )
@@ -532,7 +532,7 @@ class QueryHelper:
             create_network_security_group_details=oci_adaptor.oci.core.models.
             CreateNetworkSecurityGroupDetails(
                 compartment_id=compartment,
-                vcn_id=vcn.id,
+                vcn_id=vcn_id,
                 display_name=nsg_name,
             ))
         get_nsg_resp = net_client.get_network_security_group(

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/provision/runpod/instance.py

@@ -83,7 +83,8 @@ def run_instances(region: str, cluster_name_on_cloud: str,
         node_type = 'head' if head_instance_id is None else 'worker'
         try:
             instance_id = utils.launch(
-                name=f'{cluster_name_on_cloud}-{node_type}',
+                cluster_name=cluster_name_on_cloud,
+                node_type=node_type,
                 instance_type=config.node_config['InstanceType'],
                 region=region,
                 disk_size=config.node_config['DiskSize'],
@@ -92,6 +93,8 @@ def run_instances(region: str, cluster_name_on_cloud: str,
                 public_key=config.node_config['PublicKey'],
                 preemptible=config.node_config['Preemptible'],
                 bid_per_gpu=config.node_config['BidPerGPU'],
+                docker_login_config=config.provider_config.get(
+                    'docker_login_config'),
             )
         except Exception as e:  # pylint: disable=broad-except
             logger.warning(f'run_instances error: {e}')
@@ -145,6 +148,8 @@ def terminate_instances(
     """See sky/provision/__init__.py"""
     del provider_config  # unused
     instances = _filter_instances(cluster_name_on_cloud, None)
+    template_name, registry_auth_id = utils.get_registry_auth_resources(
+        cluster_name_on_cloud)
     for inst_id, inst in instances.items():
         logger.debug(f'Terminating instance {inst_id}: {inst}')
         if worker_only and inst['name'].endswith('-head'):
@@ -157,6 +162,10 @@ def terminate_instances(
                     f'Failed to terminate instance {inst_id}: '
                     f'{common_utils.format_exception(e, use_bracket=False)}'
                 ) from e
+    if template_name is not None:
+        utils.delete_pod_template(template_name)
+    if registry_auth_id is not None:
+        utils.delete_register_auth(registry_auth_id)
 
 
 def get_cluster_info(

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/provision/runpod/utils.py

@@ -2,10 +2,11 @@
 
 import base64
 import time
-from typing import Any, Dict, List, Optional
+from typing import Any, Dict, List, Optional, Tuple
 
 from sky import sky_logging
 from sky.adaptors import runpod
+from sky.provision import docker_utils
 import sky.provision.runpod.api.commands as runpod_commands
 from sky.skylet import constants
 from sky.utils import common_utils
@@ -47,6 +48,11 @@ GPU_NAME_MAP = {
 }
 
 
+def _construct_docker_login_template_name(cluster_name: str) -> str:
+    """Constructs the registry auth template name."""
+    return f'{cluster_name}-docker-login-template'
+
+
 def retry(func):
     """Decorator to retry a function."""
 
@@ -66,9 +72,83 @@ def retry(func):
     return wrapper
 
 
+# Adapted from runpod.api.queries.pods.py::QUERY_POD.
+# Adding containerRegistryAuthId to the query.
+_QUERY_POD = """
+query myPods {
+    myself {
+        pods {
+            id
+            containerDiskInGb
+            containerRegistryAuthId
+            costPerHr
+            desiredStatus
+            dockerArgs
+            dockerId
+            env
+            gpuCount
+            imageName
+            lastStatusChange
+            machineId
+            memoryInGb
+            name
+            podType
+            port
+            ports
+            uptimeSeconds
+            vcpuCount
+            volumeInGb
+            volumeMountPath
+            runtime {
+                ports{
+                    ip
+                    isIpPublic
+                    privatePort
+                    publicPort
+                    type
+                }
+            }
+            machine {
+                gpuDisplayName
+            }
+        }
+    }
+}
+"""
+
+
+def _sky_get_pods() -> dict:
+    """List all pods with extra registry auth information.
+
+    Adapted from runpod.get_pods() to include containerRegistryAuthId.
+    """
+    raw_return = runpod.runpod.api.graphql.run_graphql_query(_QUERY_POD)
+    cleaned_return = raw_return['data']['myself']['pods']
+    return cleaned_return
+
+
+_QUERY_POD_TEMPLATE_WITH_REGISTRY_AUTH = """
+query myself {
+    myself {
+        podTemplates {
+            name
+            containerRegistryAuthId
+        }
+    }
+}
+"""
+
+
+def _list_pod_templates_with_container_registry() -> dict:
+    """List all pod templates."""
+    raw_return = runpod.runpod.api.graphql.run_graphql_query(
+        _QUERY_POD_TEMPLATE_WITH_REGISTRY_AUTH)
+    return raw_return['data']['myself']['podTemplates']
+
+
 def list_instances() -> Dict[str, Dict[str, Any]]:
     """Lists instances associated with API key."""
-    instances = runpod.runpod.get_pods()
+    instances = _sky_get_pods()
 
     instance_dict: Dict[str, Dict[str, Any]] = {}
     for instance in instances:
@@ -100,14 +180,75 @@ def list_instances() -> Dict[str, Dict[str, Any]]:
     return instance_dict
 
 
-def launch(name: str, instance_type: str, region: str, disk_size: int,
-           image_name: str, ports: Optional[List[int]], public_key: str,
-           preemptible: Optional[bool], bid_per_gpu: float) -> str:
+def delete_pod_template(template_name: str) -> None:
+    """Deletes a pod template."""
+    try:
+        runpod.runpod.api.graphql.run_graphql_query(
+            f'mutation {{deleteTemplate(templateName: "{template_name}")}}')
+    except runpod.runpod.error.QueryError as e:
+        logger.warning(f'Failed to delete template {template_name}: {e}'
+                       'Please delete it manually.')
+
+
+def delete_register_auth(registry_auth_id: str) -> None:
+    """Deletes a registry auth."""
+    try:
+        runpod.runpod.delete_container_registry_auth(registry_auth_id)
+    except runpod.runpod.error.QueryError as e:
+        logger.warning(f'Failed to delete registry auth {registry_auth_id}: {e}'
+                       'Please delete it manually.')
+
+
+def _create_template_for_docker_login(
+    cluster_name: str,
+    image_name: str,
+    docker_login_config: Optional[Dict[str, str]],
+) -> Tuple[str, Optional[str]]:
+    """Creates a template for the given image with the docker login config.
+
+    Returns:
+        formatted_image_name: The formatted image name.
+        template_id: The template ID. None for no docker login config.
+    """
+    if docker_login_config is None:
+        return image_name, None
+    login_config = docker_utils.DockerLoginConfig(**docker_login_config)
+    container_registry_auth_name = f'{cluster_name}-registry-auth'
+    container_template_name = _construct_docker_login_template_name(
+        cluster_name)
+    # The `name` argument is only for display purpose and the registry server
+    # will be splitted from the docker image name (Tested with AWS ECR).
+    # Here we only need the username and password to create the registry auth.
+    # TODO(tian): Now we create a template and a registry auth for each cluster.
+    # Consider create one for each server and reuse them. Challenges including
+    # calculate the reference count and delete them when no longer needed.
+    create_auth_resp = runpod.runpod.create_container_registry_auth(
+        name=container_registry_auth_name,
+        username=login_config.username,
+        password=login_config.password,
+    )
+    registry_auth_id = create_auth_resp['id']
+    create_template_resp = runpod.runpod.create_template(
+        name=container_template_name,
+        image_name=None,
+        registry_auth_id=registry_auth_id,
+    )
+    return login_config.format_image(image_name), create_template_resp['id']
+
+
+def launch(cluster_name: str, node_type: str, instance_type: str, region: str,
+           disk_size: int, image_name: str, ports: Optional[List[int]],
+           public_key: str, preemptible: Optional[bool], bid_per_gpu: float,
+           docker_login_config: Optional[Dict[str, str]]) -> str:
     """Launches an instance with the given parameters.
 
     Converts the instance_type to the RunPod GPU name, finds the specs for the
     GPU, and launches the instance.
+
+    Returns:
+        instance_id: The instance ID.
     """
+    name = f'{cluster_name}-{node_type}'
     gpu_type = GPU_NAME_MAP[instance_type.split('_')[1]]
     gpu_quantity = int(instance_type.split('_')[0].replace('x', ''))
     cloud_type = instance_type.split('_')[2]
@@ -139,21 +280,24 @@ def launch(name: str, instance_type: str, region: str, disk_size: int,
     # Use base64 to deal with the tricky quoting issues caused by runpod API.
     encoded = base64.b64encode(setup_cmd.encode('utf-8')).decode('utf-8')
 
+    docker_args = (f'bash -c \'echo {encoded} | base64 --decode > init.sh; '
+                   f'bash init.sh\'')
+
     # Port 8081 is occupied for nginx in the base image.
     custom_ports_str = ''
     if ports is not None:
         custom_ports_str = ''.join([f'{p}/tcp,' for p in ports])
+    ports_str = (f'22/tcp,'
+                 f'{custom_ports_str}'
+                 f'{constants.SKY_REMOTE_RAY_DASHBOARD_PORT}/http,'
+                 f'{constants.SKY_REMOTE_RAY_PORT}/http')
 
-    docker_args = (f'bash -c \'echo {encoded} | base64 --decode > init.sh; '
-                   f'bash init.sh\'')
-    ports = (f'22/tcp,'
-             f'{custom_ports_str}'
-             f'{constants.SKY_REMOTE_RAY_DASHBOARD_PORT}/http,'
-             f'{constants.SKY_REMOTE_RAY_PORT}/http')
+    image_name_formatted, template_id = _create_template_for_docker_login(
+        cluster_name, image_name, docker_login_config)
 
     params = {
         'name': name,
-        'image_name': image_name,
+        'image_name': image_name_formatted,
         'gpu_type_id': gpu_type,
         'cloud_type': cloud_type,
         'container_disk_in_gb': disk_size,
@@ -161,9 +305,10 @@ def launch(name: str, instance_type: str, region: str, disk_size: int,
         'min_memory_in_gb': gpu_specs['memoryInGb'] * gpu_quantity,
         'gpu_count': gpu_quantity,
         'country_code': region,
-        'ports': ports,
+        'ports': ports_str,
         'support_public_ip': True,
         'docker_args': docker_args,
+        'template_id': template_id,
     }
 
     if preemptible is None or not preemptible:
@@ -177,6 +322,18 @@ def launch(name: str, instance_type: str, region: str, disk_size: int,
     return new_instance['id']
 
 
+def get_registry_auth_resources(
+        cluster_name: str) -> Tuple[Optional[str], Optional[str]]:
+    """Gets the registry auth resources."""
+    container_registry_auth_name = _construct_docker_login_template_name(
+        cluster_name)
+    for template in _list_pod_templates_with_container_registry():
+        if template['name'] == container_registry_auth_name:
+            return container_registry_auth_name, template[
+                'containerRegistryAuthId']
+    return None, None
+
+
 def remove(instance_id: str) -> None:
     """Terminates the given instance."""
     runpod.runpod.terminate_pod(instance_id)

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/setup_files/dependencies.py

@@ -123,7 +123,9 @@ extras_require: Dict[str, List[str]] = {
     'oci': ['oci'] + local_ray,
     'kubernetes': ['kubernetes>=20.0.0'],
     'remote': remote,
-    'runpod': ['runpod>=1.5.1'],
+    # For the container registry auth api. Reference:
+    # https://github.com/runpod/runpod-python/releases/tag/1.6.1
+    'runpod': ['runpod>=1.6.1'],
     'fluidstack': [],  # No dependencies needed for fluidstack
     'cudo': ['cudo-compute>=0.1.10'],
     'paperspace': [],  # No dependencies needed for paperspace

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/skylet/providers/command_runner.py

@@ -25,7 +25,7 @@ def docker_start_cmds(
     docker_cmd,
 ):
     """Generating docker start command without --rm.
-    
+
     The code is borrowed from `ray.autoscaler._private.docker`.
 
     Changes we made:
@@ -159,19 +159,17 @@ class SkyDockerCommandRunner(DockerCommandRunner):
             return True
 
         # SkyPilot: Docker login if user specified a private docker registry.
-        if "docker_login_config" in self.docker_config:
+        if 'docker_login_config' in self.docker_config:
             # TODO(tian): Maybe support a command to get the login password?
-            docker_login_config: docker_utils.DockerLoginConfig = self.docker_config[
-                "docker_login_config"]
+            docker_login_config: docker_utils.DockerLoginConfig = (
+                self.docker_config['docker_login_config'])
             self._run_with_retry(
                 f'{self.docker_cmd} login --username '
                 f'{docker_login_config.username} --password '
                 f'{docker_login_config.password} {docker_login_config.server}')
             # We automatically add the server prefix to the image name if
             # the user did not add it.
-            server_prefix = f'{docker_login_config.server}/'
-            if not specific_image.startswith(server_prefix):
-                specific_image = f'{server_prefix}{specific_image}'
+            specific_image = docker_login_config.format_image(specific_image)
 
         if self.docker_config.get('pull_before_run', True):
             assert specific_image, ('Image must be included in config if '

{skypilot_nightly-1.0.0.dev20250106 → skypilot_nightly-1.0.0.dev20250108}/sky/templates/runpod-ray.yml.j2

@@ -10,6 +10,19 @@ provider:
   module: sky.provision.runpod
   region: "{{region}}"
   disable_launch_config_check: true
+  # For RunPod, we directly set the image id for the docker as runtime environment
+  # support, thus we need to avoid the DockerInitializer detects the docker field
+  # and performs the initialization. Therefore we put the docker login config in
+  # the provider config here.
+  {%- if docker_login_config is not none %}
+  docker_login_config:
+    username: |-
+      {{docker_login_config.username}}
+    password: |-
+      {{docker_login_config.password}}
+    server: |-
+      {{docker_login_config.server}}
+  {%- endif %}
 
 auth:
   ssh_user: root