skypilot-nightly 1.0.0.dev20241020__tar.gz → 1.0.0.dev20241021__tar.gz

{skypilot_nightly-1.0.0.dev20241020/skypilot_nightly.egg-info → skypilot_nightly-1.0.0.dev20241021}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: skypilot-nightly
-Version: 1.0.0.dev20241020
+Version: 1.0.0.dev20241021
 Summary: SkyPilot: An intercloud broker for the clouds
 Author: SkyPilot Team
 License: Apache 2.0

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/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 = 'c6ae536d8dfedc3bbcf427a81480382b9d5f4c29'
+_SKYPILOT_COMMIT_SHA = '3c3bcee5cfe720a96ab67f4049a557a79e7f077f'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20241020'
+__version__ = '1.0.0.dev20241021'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/azure.py

@@ -329,7 +329,6 @@ class Azure(clouds.Cloud):
             runcmd:
               - sed -i 's/#Banner none/Banner none/' /etc/ssh/sshd_config
               - echo '\\nif [ ! -f "/tmp/__restarted" ]; then\\n  sudo systemctl restart ssh\\n  sleep 2\\n  touch /tmp/__restarted\\nfi' >> /home/skypilot:ssh_user/.bashrc
-              - usermod -aG docker skypilot:ssh_user
             write_files:
               - path: /etc/apt/apt.conf.d/20auto-upgrades
                 content: |

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/gcp.py

@@ -477,6 +477,9 @@ class GCP(clouds.Cloud):
                     'runtime_version']
                 resources_vars['tpu_node_name'] = r.accelerator_args.get(
                     'tpu_name')
+                # TPU VMs require privileged mode for docker containers to
+                # access TPU devices.
+                resources_vars['docker_run_options'] = ['--privileged']
             else:
                 # Convert to GCP names:
                 # https://cloud.google.com/compute/docs/gpus

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/lambda_cloud.py

@@ -37,10 +37,6 @@ class Lambda(clouds.Cloud):
     _CLOUD_UNSUPPORTED_FEATURES = {
         clouds.CloudImplementationFeatures.STOP: 'Lambda cloud does not support stopping VMs.',
         clouds.CloudImplementationFeatures.CLONE_DISK_FROM_CLUSTER: f'Migrating disk is currently not supported on {_REPR}.',
-        clouds.CloudImplementationFeatures.DOCKER_IMAGE: (
-            f'Docker image is currently not supported on {_REPR}. '
-            'You can try running docker command inside the `run` section in task.yaml.'
-        ),
         clouds.CloudImplementationFeatures.SPOT_INSTANCE: f'Spot instances are not supported in {_REPR}.',
         clouds.CloudImplementationFeatures.IMAGE_ID: f'Specifying image ID is not supported in {_REPR}.',
         clouds.CloudImplementationFeatures.CUSTOM_DISK_TIER: f'Custom disk tiers are not supported in {_REPR}.',
@@ -173,12 +169,20 @@ class Lambda(clouds.Cloud):
         else:
             custom_resources = None
 
-        return {
+        resources_vars = {
             'instance_type': resources.instance_type,
             'custom_resources': custom_resources,
             'region': region.name,
         }
 
+        if acc_dict is not None:
+            # Lambda cloud's docker runtime information does not contain
+            # 'nvidia-container-runtime', causing no GPU option is added to
+            # the docker run command. We patch this by adding it here.
+            resources_vars['docker_run_options'] = ['--gpus all']
+
+        return resources_vars
+
     def _get_feasible_launchable_resources(
         self, resources: 'resources_lib.Resources'
     ) -> 'resources_utils.FeasibleResources':

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/oci.py

@@ -17,6 +17,8 @@ History:
    make_deploy_resources_variables(): Bug fix for specify the image_id as
    the ocid of the image in the task.yaml file, in this case the image_id
    for the node config should be set to the ocid instead of a dict.
+ - Hysun He (hysun.he@oracle.com) @ Oct 13, 2024:
+   Support more OS types additional to ubuntu for OCI resources.
 """
 import json
 import logging
@@ -295,10 +297,21 @@ class OCI(clouds.Cloud):
             cpus=None if cpus is None else float(cpus),
             disk_tier=resources.disk_tier)
 
+        image_str = self._get_image_str(image_id=resources.image_id,
+                                        instance_type=resources.instance_type,
+                                        region=region.name)
+
+        # pylint: disable=import-outside-toplevel
+        from sky.clouds.service_catalog import oci_catalog
+        os_type = oci_catalog.get_image_os_from_tag(tag=image_str,
+                                                    region=region.name)
+        logger.debug(f'OS type for the image {image_str} is {os_type}')
+
         return {
             'instance_type': instance_type,
             'custom_resources': custom_resources,
             'region': region.name,
+            'os_type': os_type,
             'cpus': str(cpus),
             'memory': resources.memory,
             'disk_size': resources.disk_size,
@@ -501,59 +514,45 @@ class OCI(clouds.Cloud):
         region_name: str,
         instance_type: str,
     ) -> str:
-        if image_id is None:
-            return self._get_default_image(region_name=region_name,
-                                           instance_type=instance_type)
-        if None in image_id:
-            image_id_str = image_id[None]
-        else:
-            assert region_name in image_id, image_id
-            image_id_str = image_id[region_name]
+        image_id_str = self._get_image_str(image_id=image_id,
+                                           instance_type=instance_type,
+                                           region=region_name)
+
         if image_id_str.startswith('skypilot:'):
             image_id_str = service_catalog.get_image_id_from_tag(image_id_str,
                                                                  region_name,
                                                                  clouds='oci')
-            if image_id_str is None:
-                logger.critical(
-                    '! Real image_id not found! - {region_name}:{image_id}')
-                # Raise ResourcesUnavailableError to make sure the failover
-                # in CloudVMRayBackend will be correctly triggered.
-                # TODO(zhwu): This is a information leakage to the cloud
-                # implementor, we need to find a better way to handle this.
-                raise exceptions.ResourcesUnavailableError(
-                    '! ERR: No image found in catalog for region '
-                    f'{region_name}. Try setting a valid image_id.')
+
+        # Image_id should be impossible be None, except for the case when
+        # user specify an image tag which does not exist in the image.csv
+        # catalog file which only possible in "test" / "evaluation" phase.
+        # Therefore, we use assert here.
+        assert image_id_str is not None
 
         logger.debug(f'Got real image_id {image_id_str}')
         return image_id_str
 
-    def _get_default_image(self, region_name: str, instance_type: str) -> str:
+    def _get_image_str(self, image_id: Optional[Dict[Optional[str], str]],
+                       instance_type: str, region: str):
+        if image_id is None:
+            image_str = self._get_default_image_tag(instance_type)
+        elif None in image_id:
+            image_str = image_id[None]
+        else:
+            assert region in image_id, image_id
+            image_str = image_id[region]
+        return image_str
+
+    def _get_default_image_tag(self, instance_type: str) -> str:
         acc = self.get_accelerators_from_instance_type(instance_type)
 
         if acc is None:
             image_tag = oci_utils.oci_config.get_default_image_tag()
-            image_id_str = service_catalog.get_image_id_from_tag(image_tag,
-                                                                 region_name,
-                                                                 clouds='oci')
         else:
             assert len(acc) == 1, acc
             image_tag = oci_utils.oci_config.get_default_gpu_image_tag()
-            image_id_str = service_catalog.get_image_id_from_tag(image_tag,
-                                                                 region_name,
-                                                                 clouds='oci')
 
-        if image_id_str is not None:
-            logger.debug(
-                f'Got default image_id {image_id_str} from tag {image_tag}')
-            return image_id_str
-
-        # Raise ResourcesUnavailableError to make sure the failover in
-        # CloudVMRayBackend will be correctly triggered.
-        # TODO(zhwu): This is a information leakage to the cloud implementor,
-        # we need to find a better way to handle this.
-        raise exceptions.ResourcesUnavailableError(
-            'ERR: No image found in catalog for region '
-            f'{region_name}. Try update your default image_id settings.')
+        return image_tag
 
     def get_vpu_from_disktier(
             self, cpus: Optional[float],

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/service_catalog/oci_catalog.py

@@ -7,6 +7,8 @@ History:
  - Hysun He (hysun.he@oracle.com) @ Apr, 2023: Initial implementation
  - Hysun He (hysun.he@oracle.com) @ Jun, 2023: Reduce retry times by
    excluding those unsubscribed regions.
+ - Hysun He (hysun.he@oracle.com) @ Oct 14, 2024: Bug fix for validation
+   of the Marketplace images
 """
 
 import logging
@@ -206,4 +208,24 @@ def get_image_id_from_tag(tag: str, region: Optional[str]) -> Optional[str]:
 
 def is_image_tag_valid(tag: str, region: Optional[str]) -> bool:
     """Returns whether the image tag is valid."""
+    # Oct.14, 2024 by Hysun He: Marketplace images are region neutral, so don't
+    # check with region for the Marketplace images.
+    df = _image_df[_image_df['Tag'].str.fullmatch(tag)]
+    if df.empty:
+        return False
+    app_catalog_listing_id = df['AppCatalogListingId'].iloc[0]
+    if app_catalog_listing_id:
+        return True
     return common.is_image_tag_valid_impl(_image_df, tag, region)
+
+
+def get_image_os_from_tag(tag: str, region: Optional[str]) -> Optional[str]:
+    del region
+    df = _image_df[_image_df['Tag'].str.fullmatch(tag)]
+    if df.empty:
+        os_type = oci_utils.oci_config.get_default_image_os()
+    else:
+        os_type = df['OS'].iloc[0]
+
+    logger.debug(f'Operation system for the image {tag} is {os_type}')
+    return os_type

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/clouds/utils/oci_utils.py

@@ -1,7 +1,9 @@
 """OCI Configuration.
 History:
- - Zhanghao Wu @ Oct 2023: Formatting and refactoring
  - Hysun He (hysun.he@oracle.com) @ Apr, 2023: Initial implementation
+ - Zhanghao Wu @ Oct 2023: Formatting and refactoring
+ - Hysun He (hysun.he@oracle.com) @ Oct, 2024: Add default image OS
+   configuration.
 """
 import logging
 import os
@@ -121,5 +123,13 @@ class OCIConfig:
         return skypilot_config.get_nested(
             ('oci', 'default', 'oci_config_profile'), 'DEFAULT')
 
+    @classmethod
+    def get_default_image_os(cls) -> str:
+        # Get the default image OS. Instead of hardcoding, we give a choice to
+        # set the default image OS type in the sky's user-config file. (if not
+        # specified, use the hardcode one at last)
+        return skypilot_config.get_nested(('oci', 'default', 'image_os_type'),
+                                          'ubuntu')
+
 
 oci_config = OCIConfig()

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/provision/docker_utils.py

@@ -253,12 +253,13 @@ class DockerInitializer:
             # issue with nvidia container toolkit:
             # https://github.com/NVIDIA/nvidia-container-toolkit/issues/48
             self._run(
-                '[ -f /etc/docker/daemon.json ] || '
+                '{ which jq || sudo apt update && sudo apt install -y jq; } && '
+                '{ [ -f /etc/docker/daemon.json ] || '
                 'echo "{}" | sudo tee /etc/docker/daemon.json;'
                 'sudo jq \'.["exec-opts"] = ["native.cgroupdriver=cgroupfs"]\' '
                 '/etc/docker/daemon.json > /tmp/daemon.json;'
                 'sudo mv /tmp/daemon.json /etc/docker/daemon.json;'
-                'sudo systemctl restart docker')
+                'sudo systemctl restart docker; } || true')
             user_docker_run_options = self.docker_config.get('run_options', [])
             start_command = docker_start_cmds(
                 specific_image,
@@ -335,7 +336,11 @@ class DockerInitializer:
 
     def _check_docker_installed(self):
         no_exist = 'NoExist'
+        # SkyPilot: Add the current user to the docker group first (if needed),
+        # before checking if docker is installed to avoid permission issues.
         cleaned_output = self._run(
+            'id -nG $USER | grep -qw docker || '
+            'sudo usermod -aG docker $USER > /dev/null 2>&1;'
             f'command -v {self.docker_cmd} || echo {no_exist!r}')
         if no_exist in cleaned_output or 'docker' not in cleaned_output:
             logger.error(
@@ -424,8 +429,8 @@ class DockerInitializer:
     def _check_container_exited(self) -> bool:
         if self.initialized:
             return True
-        output = (self._run(check_docker_running_cmd(self.container_name,
-                                                     self.docker_cmd),
-                            wait_for_docker_daemon=True))
-        return 'false' in output.lower(
-        ) and 'no such object' not in output.lower()
+        output = self._run(check_docker_running_cmd(self.container_name,
+                                                    self.docker_cmd),
+                           wait_for_docker_daemon=True)
+        return ('false' in output.lower() and
+                'no such object' not in output.lower())

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/provision/paperspace/utils.py

@@ -132,6 +132,8 @@ class PaperspaceCloudClient:
             'apt-get update \n'
             'apt-get install -y docker-ce docker-ce-cli containerd.io docker-buildx-plugin docker-compose-plugin \n'  # pylint: disable=line-too-long
             'fi \n'
+            # TODO(tian): Maybe remove this as well since we are now adding
+            # users to docker group in the DockerInitializer. Need to test.
             'usermod -aG docker paperspace \n'
             f'echo "{public_key}" >> /home/paperspace/.ssh/authorized_keys \n')
         try:

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/resources.py

@@ -842,12 +842,6 @@ class Resources:
 
         if self.extract_docker_image() is not None:
             # TODO(tian): validate the docker image exists / of reasonable size
-            if self.accelerators is not None:
-                for acc in self.accelerators.keys():
-                    if acc.lower().startswith('tpu'):
-                        with ux_utils.print_exception_no_traceback():
-                            raise ValueError(
-                                'Docker image is not supported for TPU VM.')
             if self.cloud is not None:
                 self.cloud.check_features_are_supported(
                     self, {clouds.CloudImplementationFeatures.DOCKER_IMAGE})
@@ -1032,6 +1026,12 @@ class Resources:
                 self.accelerators is not None):
             initial_setup_commands = [constants.DISABLE_GPU_ECC_COMMAND]
 
+        docker_image = self.extract_docker_image()
+
+        # Cloud specific variables
+        cloud_specific_variables = self.cloud.make_deploy_resources_variables(
+            self, cluster_name, region, zones, dryrun)
+
         # Docker run options
         docker_run_options = skypilot_config.get_nested(
             ('docker', 'run_options'),
@@ -1039,18 +1039,17 @@ class Resources:
             override_configs=self.cluster_config_overrides)
         if isinstance(docker_run_options, str):
             docker_run_options = [docker_run_options]
+        # Special accelerator runtime might require additional docker run
+        # options. e.g., for TPU, we need --privileged.
+        if 'docker_run_options' in cloud_specific_variables:
+            docker_run_options.extend(
+                cloud_specific_variables['docker_run_options'])
         if docker_run_options and isinstance(self.cloud, clouds.Kubernetes):
             logger.warning(
                 f'{colorama.Style.DIM}Docker run options are specified, '
                 'but ignored for Kubernetes: '
                 f'{" ".join(docker_run_options)}'
                 f'{colorama.Style.RESET_ALL}')
-
-        docker_image = self.extract_docker_image()
-
-        # Cloud specific variables
-        cloud_specific_variables = self.cloud.make_deploy_resources_variables(
-            self, cluster_name, region, zones, dryrun)
         return dict(
             cloud_specific_variables,
             **{

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/serve/serve_utils.py

@@ -246,9 +246,11 @@ def set_service_status_and_active_versions_from_replica(
         update_mode: UpdateMode) -> None:
     record = serve_state.get_service_from_name(service_name)
     if record is None:
-        raise ValueError('The service is up-ed in an old version and does not '
-                         'support update. Please `sky serve down` '
-                         'it first and relaunch the service.')
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(
+                'The service is up-ed in an old version and does not '
+                'support update. Please `sky serve down` '
+                'it first and relaunch the service.')
     if record['status'] == serve_state.ServiceStatus.SHUTTING_DOWN:
         # When the service is shutting down, there is a period of time which the
         # controller still responds to the request, and the replica is not
@@ -289,7 +291,8 @@ def update_service_status() -> None:
 def update_service_encoded(service_name: str, version: int, mode: str) -> str:
     service_status = _get_service_status(service_name)
     if service_status is None:
-        raise ValueError(f'Service {service_name!r} does not exist.')
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(f'Service {service_name!r} does not exist.')
     controller_port = service_status['controller_port']
     resp = requests.post(
         _CONTROLLER_URL.format(CONTROLLER_PORT=controller_port) +
@@ -299,15 +302,21 @@ def update_service_encoded(service_name: str, version: int, mode: str) -> str:
             'mode': mode,
         })
     if resp.status_code == 404:
-        raise ValueError('The service is up-ed in an old version and does not '
-                         'support update. Please `sky serve down` '
-                         'it first and relaunch the service. ')
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(
+                'The service is up-ed in an old version and does not '
+                'support update. Please `sky serve down` '
+                'it first and relaunch the service. ')
     elif resp.status_code == 400:
-        raise ValueError(f'Client error during service update: {resp.text}')
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(f'Client error during service update: {resp.text}')
     elif resp.status_code == 500:
-        raise RuntimeError(f'Server error during service update: {resp.text}')
+        with ux_utils.print_exception_no_traceback():
+            raise RuntimeError(
+                f'Server error during service update: {resp.text}')
     elif resp.status_code != 200:
-        raise ValueError(f'Failed to update service: {resp.text}')
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(f'Failed to update service: {resp.text}')
 
     service_msg = resp.json()['message']
     return common_utils.encode_payload(service_msg)
@@ -674,8 +683,9 @@ def stream_replica_logs(service_name: str, replica_id: int,
         for info in replica_info:
             if info.replica_id == replica_id:
                 return info.status
-        raise ValueError(
-            _FAILED_TO_FIND_REPLICA_MSG.format(replica_id=replica_id))
+        with ux_utils.print_exception_no_traceback():
+            raise ValueError(
+                _FAILED_TO_FIND_REPLICA_MSG.format(replica_id=replica_id))
 
     finish_stream = (
         lambda: _get_replica_status() != serve_state.ReplicaStatus.PROVISIONING)

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/templates/lambda-ray.yml.j2

@@ -5,6 +5,26 @@ max_workers: {{num_nodes - 1}}
 upscaling_speed: {{num_nodes - 1}}
 idle_timeout_minutes: 60
 
+{%- if docker_image is not none %}
+docker:
+  image: {{docker_image}}
+  container_name: {{docker_container_name}}
+  run_options:
+    - --ulimit nofile=1048576:1048576
+    {%- for run_option in docker_run_options %}
+    - {{run_option}}
+    {%- endfor %}
+  {%- 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 %}
+{%- endif %}
+
 provider:
   type: external
   module: sky.provision.lambda

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/templates/oci-ray.yml.j2

@@ -16,7 +16,11 @@ provider:
   disable_launch_config_check: true
 
 auth:
+{% if os_type == "ubuntu" %}
   ssh_user: ubuntu
+{% else %}
+  ssh_user: opc
+{% endif %}
   ssh_private_key: {{ssh_private_key}}
 
 available_node_types:
@@ -85,14 +89,20 @@ setup_commands:
   # Line 'sudo grep ..': set the number of threads per process to unlimited to avoid ray job submit stucking issue when the number of running ray jobs increase.
   # Line 'mkdir -p ..': disable host key check
   # Line 'python3 -c ..': patch the buggy ray files and enable `-o allow_other` option for `goofys`
-  - sudo systemctl stop unattended-upgrades || true;
+  - echo "setup commands runs at $(date)" > /tmp/provision.tmp.out || true;
+    {%- if os_type == "ubuntu" %}
+    sudo systemctl stop unattended-upgrades || true;
     sudo systemctl disable unattended-upgrades || true;
     sudo sed -i 's/Unattended-Upgrade "1"/Unattended-Upgrade "0"/g' /etc/apt/apt.conf.d/20auto-upgrades || true;
     sudo kill -9 `sudo lsof /var/lib/dpkg/lock-frontend | awk '{print $2}' | tail -n 1` || true;
     sudo pkill -9 apt-get;
     sudo pkill -9 dpkg;
     sudo dpkg --configure -a;
-    ([ `sudo lshw -class display | grep "NVIDIA Corporation" | wc -l` -gt 0 ]) && (sudo which nvidia-smi > /dev/null || ( sudo apt-get install nvidia-driver-530-open -y && sudo apt-get install nvidia-driver-525-server -y ) || true);
+    {%- else %}
+    sudo /usr/libexec/oci-growfs -y || true;
+    sudo systemctl stop firewalld || true;
+    sudo systemctl disable firewalld || true;
+    {%- endif %}
     mkdir -p ~/.ssh; touch ~/.ssh/config;
     {{ conda_installation_commands }}
     {{ ray_skypilot_installation_commands }}

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021}/sky/utils/command_runner.py

@@ -502,8 +502,10 @@ class SSHCommandRunner(CommandRunner):
         if self.ssh_control_name is not None:
             control_path = _ssh_control_path(self.ssh_control_name)
             if control_path is not None:
+                # Suppress the `Exit request sent.` output for this comamnd
+                # which would interrupt the CLI spinner.
                 cmd = (f'ssh -O exit -S {control_path}/%C '
-                       f'{self.ssh_user}@{self.ip}')
+                       f'{self.ssh_user}@{self.ip} > /dev/null 2>&1')
                 logger.debug(f'Closing cached connection {control_path!r} with '
                              f'cmd: {cmd}')
                 log_lib.run_with_log(cmd,

{skypilot_nightly-1.0.0.dev20241020 → skypilot_nightly-1.0.0.dev20241021/skypilot_nightly.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: skypilot-nightly
-Version: 1.0.0.dev20241020
+Version: 1.0.0.dev20241021
 Summary: SkyPilot: An intercloud broker for the clouds
 Author: SkyPilot Team
 License: Apache 2.0