skypilot-nightly 1.0.0.dev20241110__tar.gz → 1.0.0.dev20241112__tar.gz

{skypilot_nightly-1.0.0.dev20241110/skypilot_nightly.egg-info → skypilot_nightly-1.0.0.dev20241112}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: skypilot-nightly
-Version: 1.0.0.dev20241110
+Version: 1.0.0.dev20241112
 Summary: SkyPilot: An intercloud broker for the clouds
 Author: SkyPilot Team
 License: Apache 2.0
@@ -319,7 +319,7 @@ Runnable examples:
   - [LocalGPT](./llm/localgpt)
   - [Falcon](./llm/falcon)
   - Add yours here & see more in [`llm/`](./llm)!
-- Framework examples: [PyTorch DDP](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_torch.yaml), [DeepSpeed](./examples/deepspeed-multinode/sky.yaml), [JAX/Flax on TPU](https://github.com/skypilot-org/skypilot/blob/master/examples/tpu/tpuvm_mnist.yaml), [Stable Diffusion](https://github.com/skypilot-org/skypilot/tree/master/examples/stable_diffusion), [Detectron2](https://github.com/skypilot-org/skypilot/blob/master/examples/detectron2_docker.yaml), [Distributed](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_tf_app.py) [TensorFlow](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_app_storage.yaml), [Ray Train](examples/distributed_ray_train/ray_train.yaml), [NeMo](https://github.com/skypilot-org/skypilot/blob/master/examples/nemo/nemo.yaml), [programmatic grid search](https://github.com/skypilot-org/skypilot/blob/master/examples/huggingface_glue_imdb_grid_search_app.py), [Docker](https://github.com/skypilot-org/skypilot/blob/master/examples/docker/echo_app.yaml), [Cog](https://github.com/skypilot-org/skypilot/blob/master/examples/cog/), [Unsloth](https://github.com/skypilot-org/skypilot/blob/master/examples/unsloth/unsloth.yaml), [Ollama](https://github.com/skypilot-org/skypilot/blob/master/llm/ollama), [llm.c](https://github.com/skypilot-org/skypilot/tree/master/llm/gpt-2), [Airflow](./examples/airflow/training_workflow) and [many more (`examples/`)](./examples).
+- Framework examples: [PyTorch DDP](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_torch.yaml), [DeepSpeed](./examples/deepspeed-multinode/sky.yaml), [JAX/Flax on TPU](https://github.com/skypilot-org/skypilot/blob/master/examples/tpu/tpuvm_mnist.yaml), [Stable Diffusion](https://github.com/skypilot-org/skypilot/tree/master/examples/stable_diffusion), [Detectron2](https://github.com/skypilot-org/skypilot/blob/master/examples/detectron2_docker.yaml), [Distributed](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_tf_app.py) [TensorFlow](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_app_storage.yaml), [Ray Train](examples/distributed_ray_train/ray_train.yaml), [NeMo](https://github.com/skypilot-org/skypilot/blob/master/examples/nemo/), [programmatic grid search](https://github.com/skypilot-org/skypilot/blob/master/examples/huggingface_glue_imdb_grid_search_app.py), [Docker](https://github.com/skypilot-org/skypilot/blob/master/examples/docker/echo_app.yaml), [Cog](https://github.com/skypilot-org/skypilot/blob/master/examples/cog/), [Unsloth](https://github.com/skypilot-org/skypilot/blob/master/examples/unsloth/unsloth.yaml), [Ollama](https://github.com/skypilot-org/skypilot/blob/master/llm/ollama), [llm.c](https://github.com/skypilot-org/skypilot/tree/master/llm/gpt-2), [Airflow](./examples/airflow/training_workflow) and [many more (`examples/`)](./examples).
 
 Case Studies and Integrations: [Community Spotlights](https://blog.skypilot.co/community/)
 

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/README.md

@@ -183,7 +183,7 @@ Runnable examples:
   - [LocalGPT](./llm/localgpt)
   - [Falcon](./llm/falcon)
   - Add yours here & see more in [`llm/`](./llm)!
-- Framework examples: [PyTorch DDP](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_torch.yaml), [DeepSpeed](./examples/deepspeed-multinode/sky.yaml), [JAX/Flax on TPU](https://github.com/skypilot-org/skypilot/blob/master/examples/tpu/tpuvm_mnist.yaml), [Stable Diffusion](https://github.com/skypilot-org/skypilot/tree/master/examples/stable_diffusion), [Detectron2](https://github.com/skypilot-org/skypilot/blob/master/examples/detectron2_docker.yaml), [Distributed](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_tf_app.py) [TensorFlow](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_app_storage.yaml), [Ray Train](examples/distributed_ray_train/ray_train.yaml), [NeMo](https://github.com/skypilot-org/skypilot/blob/master/examples/nemo/nemo.yaml), [programmatic grid search](https://github.com/skypilot-org/skypilot/blob/master/examples/huggingface_glue_imdb_grid_search_app.py), [Docker](https://github.com/skypilot-org/skypilot/blob/master/examples/docker/echo_app.yaml), [Cog](https://github.com/skypilot-org/skypilot/blob/master/examples/cog/), [Unsloth](https://github.com/skypilot-org/skypilot/blob/master/examples/unsloth/unsloth.yaml), [Ollama](https://github.com/skypilot-org/skypilot/blob/master/llm/ollama), [llm.c](https://github.com/skypilot-org/skypilot/tree/master/llm/gpt-2), [Airflow](./examples/airflow/training_workflow) and [many more (`examples/`)](./examples).
+- Framework examples: [PyTorch DDP](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_torch.yaml), [DeepSpeed](./examples/deepspeed-multinode/sky.yaml), [JAX/Flax on TPU](https://github.com/skypilot-org/skypilot/blob/master/examples/tpu/tpuvm_mnist.yaml), [Stable Diffusion](https://github.com/skypilot-org/skypilot/tree/master/examples/stable_diffusion), [Detectron2](https://github.com/skypilot-org/skypilot/blob/master/examples/detectron2_docker.yaml), [Distributed](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_distributed_tf_app.py) [TensorFlow](https://github.com/skypilot-org/skypilot/blob/master/examples/resnet_app_storage.yaml), [Ray Train](examples/distributed_ray_train/ray_train.yaml), [NeMo](https://github.com/skypilot-org/skypilot/blob/master/examples/nemo/), [programmatic grid search](https://github.com/skypilot-org/skypilot/blob/master/examples/huggingface_glue_imdb_grid_search_app.py), [Docker](https://github.com/skypilot-org/skypilot/blob/master/examples/docker/echo_app.yaml), [Cog](https://github.com/skypilot-org/skypilot/blob/master/examples/cog/), [Unsloth](https://github.com/skypilot-org/skypilot/blob/master/examples/unsloth/unsloth.yaml), [Ollama](https://github.com/skypilot-org/skypilot/blob/master/llm/ollama), [llm.c](https://github.com/skypilot-org/skypilot/tree/master/llm/gpt-2), [Airflow](./examples/airflow/training_workflow) and [many more (`examples/`)](./examples).
 
 Case Studies and Integrations: [Community Spotlights](https://blog.skypilot.co/community/)
 

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/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 = 'dddd65187953a5d6b32f762bea78eed1f109ec3c'
+_SKYPILOT_COMMIT_SHA = '140125eaad5fb64da37934c8f6650d68aa135f77'
 
 
 def _get_git_commit():
@@ -35,7 +35,7 @@ def _get_git_commit():
 
 
 __commit__ = _get_git_commit()
-__version__ = '1.0.0.dev20241110'
+__version__ = '1.0.0.dev20241112'
 __root_dir__ = os.path.dirname(os.path.abspath(__file__))
 
 

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/cli.py

@@ -3102,6 +3102,7 @@ def show_gpus(
     kubernetes_autoscaling = kubernetes_utils.get_autoscaler_type() is not None
     kubernetes_is_enabled = sky_clouds.cloud_in_iterable(
         sky_clouds.Kubernetes(), global_user_state.get_cached_enabled_clouds())
+    no_permissions_str = '<no permissions>'
 
     def _list_to_str(lst):
         return ', '.join([str(e) for e in lst])
@@ -3146,9 +3147,11 @@ def show_gpus(
                             debug_msg)
             raise ValueError(full_err_msg)
         for gpu, _ in sorted(counts.items()):
+            available_qty = available[gpu] if available[gpu] != -1 else (
+                no_permissions_str)
             realtime_gpu_table.add_row([
                 gpu,
-                _list_to_str(counts.pop(gpu)), capacity[gpu], available[gpu]
+                _list_to_str(counts.pop(gpu)), capacity[gpu], available_qty
             ])
         return realtime_gpu_table
 
@@ -3158,10 +3161,11 @@ def show_gpus(
 
         node_info_dict = kubernetes_utils.get_kubernetes_node_info(context)
         for node_name, node_info in node_info_dict.items():
+            available = node_info.free['nvidia.com/gpu'] if node_info.free[
+                'nvidia.com/gpu'] != -1 else no_permissions_str
             node_table.add_row([
                 node_name, node_info.gpu_type,
-                node_info.total['nvidia.com/gpu'],
-                node_info.free['nvidia.com/gpu']
+                node_info.total['nvidia.com/gpu'], available
             ])
         return node_table
 

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/clouds/service_catalog/kubernetes_catalog.py

@@ -10,6 +10,7 @@ from typing import Dict, List, Optional, Set, Tuple
 from sky import check as sky_check
 from sky import sky_logging
 from sky.adaptors import common as adaptors_common
+from sky.adaptors import kubernetes
 from sky.clouds import Kubernetes
 from sky.clouds.service_catalog import CloudFilter
 from sky.clouds.service_catalog import common
@@ -22,6 +23,8 @@ if typing.TYPE_CHECKING:
 else:
     pd = adaptors_common.LazyImport('pandas')
 
+logger = sky_logging.init_logger(__name__)
+
 _PULL_FREQUENCY_HOURS = 7
 
 # We keep pull_frequency_hours so we can remotely update the default image paths
@@ -77,6 +80,11 @@ def list_accelerators_realtime(
     require_price: bool = True
 ) -> Tuple[Dict[str, List[common.InstanceTypeInfo]], Dict[str, int], Dict[str,
                                                                           int]]:
+    """List accelerators in the Kubernetes cluster.
+
+    If the user does not have sufficient permissions to list pods in all
+    namespaces, the function will return free GPUs as -1.
+    """
     # TODO(romilb): This should be refactored to use get_kubernetes_node_info()
     #   function from kubernetes_utils.
     del all_regions, require_price  # Unused.
@@ -108,7 +116,17 @@ def list_accelerators_realtime(
     key = label_formatter.get_label_key()
     nodes = kubernetes_utils.get_kubernetes_nodes(context)
     # Get the pods to get the real-time GPU usage
-    pods = kubernetes_utils.get_all_pods_in_kubernetes_cluster(context)
+    try:
+        pods = kubernetes_utils.get_all_pods_in_kubernetes_cluster(context)
+    except kubernetes.api_exception() as e:
+        if e.status == 403:
+            logger.warning('Failed to get pods in the Kubernetes cluster '
+                           '(forbidden). Please check if your account has '
+                           'necessary permissions to list pods. Realtime GPU '
+                           'availability information may be incorrect.')
+            pods = None
+        else:
+            raise
     # Total number of GPUs in the cluster
     total_accelerators_capacity: Dict[str, int] = {}
     # Total number of GPUs currently available in the cluster
@@ -141,6 +159,21 @@ def list_accelerators_realtime(
                 if accelerator_count not in accelerators_qtys:
                     accelerators_qtys.add((accelerator_name, accelerator_count))
 
+            if accelerator_count >= min_quantity_filter:
+                quantized_count = (min_quantity_filter *
+                                   (accelerator_count // min_quantity_filter))
+                if accelerator_name not in total_accelerators_capacity:
+                    total_accelerators_capacity[
+                        accelerator_name] = quantized_count
+                else:
+                    total_accelerators_capacity[
+                        accelerator_name] += quantized_count
+
+            if pods is None:
+                # If we can't get the pods, we can't get the GPU usage
+                total_accelerators_available[accelerator_name] = -1
+                continue
+
             for pod in pods:
                 # Get all the pods running on the node
                 if (pod.spec.node_name == node.metadata.name and
@@ -155,16 +188,6 @@ def list_accelerators_realtime(
 
             accelerators_available = accelerator_count - allocated_qty
 
-            if accelerator_count >= min_quantity_filter:
-                quantized_count = (min_quantity_filter *
-                                   (accelerator_count // min_quantity_filter))
-                if accelerator_name not in total_accelerators_capacity:
-                    total_accelerators_capacity[
-                        accelerator_name] = quantized_count
-                else:
-                    total_accelerators_capacity[
-                        accelerator_name] += quantized_count
-
             if accelerator_name not in total_accelerators_available:
                 total_accelerators_available[accelerator_name] = 0
             if accelerators_available >= min_quantity_filter:

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/provision/docker_utils.py

@@ -20,7 +20,7 @@ SETUP_ENV_VARS_CMD = (
     '{ if [ $(id -u) -ne 0 ]; then echo "sudo"; else echo ""; fi; } && '
     'printenv | while IFS=\'=\' read -r key value; do echo "export $key=\\\"$value\\\""; done > '  # pylint: disable=line-too-long
     '~/container_env_var.sh && '
-    '$(prefix_cmd) mv ~/container_env_var.sh /etc/profile.d/container_env_var.sh'
+    '$(prefix_cmd) mv ~/container_env_var.sh /etc/profile.d/container_env_var.sh;'
 )
 
 # Docker daemon may not be ready when the machine is firstly started. The error

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/provision/kubernetes/instance.py

@@ -333,52 +333,37 @@ def _run_function_with_retries(func: Callable,
                 raise
 
 
-def _set_env_vars_in_pods(namespace: str, context: Optional[str],
-                          new_pods: List):
-    """Setting environment variables in pods.
-
-    Once all containers are ready, we can exec into them and set env vars.
-    Kubernetes automatically populates containers with critical
-    environment variables, such as those for discovering services running
-    in the cluster and CUDA/nvidia environment variables. We need to
-    make sure these env vars are available in every task and ssh session.
-    This is needed for GPU support and service discovery.
-    See https://github.com/skypilot-org/skypilot/issues/2287 for
-    more details.
-
-    To do so, we capture env vars from the pod's runtime and write them to
-    /etc/profile.d/, making them available for all users in future
-    shell sessions.
-    """
-    set_k8s_env_var_cmd = docker_utils.SETUP_ENV_VARS_CMD
+def pre_init(namespace: str, context: Optional[str], new_nodes: List) -> None:
+    """Pre-initialization step for SkyPilot pods.
 
-    def _set_env_vars_thread(new_pod):
-        pod_name = new_pod.metadata.name
-        logger.info(f'{"-"*20}Start: Set up env vars in pod {pod_name!r} '
-                    f'{"-"*20}')
-        runner = command_runner.KubernetesCommandRunner(
-            ((namespace, context), pod_name))
+    This step is run in the pod right after it is created and before the
+    SkyPilot runtime is setup.
 
-        def _run_env_vars_cmd():
-            rc, stdout, _ = runner.run(set_k8s_env_var_cmd,
-                                       require_outputs=True,
-                                       stream_logs=False)
-            _raise_command_running_error('set env vars', set_k8s_env_var_cmd,
-                                         pod_name, rc, stdout)
+    This step includes three key steps:
 
-        _run_function_with_retries(_run_env_vars_cmd,
-                                   f'set env vars in pod {pod_name}')
-        logger.info(f'{"-"*20}End: Set up env vars in pod {pod_name!r} '
-                    f'{"-"*20}')
+    1. Privilege check: Checks if the default user has sufficient privilege
+    to set up the kubernetes instance pod.
+    2. SSH setup: Sets up SSH for the pod instance.
+    3. Environment variable setup to populate k8s env vars in the pod.
 
-    subprocess_utils.run_in_parallel(_set_env_vars_thread, new_pods,
-                                     NUM_THREADS)
+    Make sure commands used in these methods are generic and work
+    on most base images. E.g., do not use Python, since that may not
+    be installed by default.
 
+    If you run any apt commands, be sure to check if the lock is available.
+    It is possible the `apt update` run in the pod container args may still
+    be running.
+
+    Args:
+        namespace (str): Kubernetes namespace.
+        context (Optional[str]): Kubernetes context.
+        new_nodes (List): List of new pod instances.
+
+    Raises:
+        config_lib.KubernetesError: If user privileges are insufficient or
+          setup fails.
+    """
 
-def _check_user_privilege(namespace: str, context: Optional[str],
-                          new_nodes: List) -> None:
-    # Checks if the default user has sufficient privilege to set up
-    # the kubernetes instance pod.
     check_k8s_user_sudo_cmd = (
         'if [ $(id -u) -eq 0 ]; then'
         # If user is root, create an alias for sudo used in skypilot setup
@@ -386,56 +371,67 @@ def _check_user_privilege(namespace: str, context: Optional[str],
         'else '
         '  if command -v sudo >/dev/null 2>&1; then '
         '    timeout 2 sudo -l >/dev/null 2>&1 && echo succeed || '
-        f'    ( echo {exceptions.INSUFFICIENT_PRIVILEGES_CODE!r}; ); '
+        f'    ( echo {exceptions.INSUFFICIENT_PRIVILEGES_CODE!r}; '
+        f'      exit {exceptions.INSUFFICIENT_PRIVILEGES_CODE}; ); '
         '  else '
-        f'    ( echo {exceptions.INSUFFICIENT_PRIVILEGES_CODE!r}; ); '
+        f'    ( echo {exceptions.INSUFFICIENT_PRIVILEGES_CODE!r}; '
+        f'      exit {exceptions.INSUFFICIENT_PRIVILEGES_CODE}; ); '
         '  fi; '
-        'fi')
+        'fi;')
+
+    # Kubernetes automatically populates containers with critical
+    # environment variables, such as those for discovering services running
+    # in the cluster and CUDA/nvidia environment variables. We need to
+    # make sure these env vars are available in every task and ssh session.
+    # This is needed for GPU support and service discovery.
+    # See https://github.com/skypilot-org/skypilot/issues/2287 for more details.
+    # To do so, we capture env vars from the pod's runtime and write them to
+    # /etc/profile.d/, making them available for all users in future
+    # shell sessions.
+    set_k8s_env_var_cmd = docker_utils.SETUP_ENV_VARS_CMD
 
-    # This check needs to run on a per-image basis, so running the check on
-    # any one pod is sufficient.
-    new_node = new_nodes[0]
-    pod_name = new_node.metadata.name
+    check_apt_update_complete_cmd = (
+        'echo "Checking if apt update from container init is complete..."; '
+        'timeout_secs=600; '
+        'start_time=$(date +%s); '
+        'while ! grep -q "Fetched" /tmp/apt-update.log 2>/dev/null; do '
+        '  echo "apt update still running. Logs:"; '
+        '  cat /tmp/apt-update.log; '
+        '  current_time=$(date +%s); '
+        '  elapsed=$((current_time - start_time)); '
+        '  if [ $elapsed -ge $timeout_secs ]; then '
+        '    echo "Timed out waiting for apt update"; '
+        '    exit 1; '
+        '  fi; '
+        '  sleep 5; '
+        'done; '
+        'echo "apt update complete."; ')
 
-    runner = command_runner.KubernetesCommandRunner(
-        ((namespace, context), pod_name))
-    logger.info(f'{"-"*20}Start: Check user privilege in pod {pod_name!r} '
-                f'{"-"*20}')
-
-    def _run_privilege_check():
-        rc, stdout, stderr = runner.run(check_k8s_user_sudo_cmd,
-                                        require_outputs=True,
-                                        separate_stderr=True,
-                                        stream_logs=False)
-        _raise_command_running_error('check user privilege',
-                                     check_k8s_user_sudo_cmd, pod_name, rc,
-                                     stdout + stderr)
-        return stdout
-
-    stdout = _run_function_with_retries(
-        _run_privilege_check, f'check user privilege in pod {pod_name!r}')
-
-    if stdout == str(exceptions.INSUFFICIENT_PRIVILEGES_CODE):
-        raise config_lib.KubernetesError(
-            'Insufficient system privileges detected. '
-            'Ensure the default user has root access or '
-            '"sudo" is installed and the user is added to the sudoers '
-            'from the image.')
-    logger.info(f'{"-"*20}End: Check user privilege in pod {pod_name!r} '
-                f'{"-"*20}')
-
-
-def _setup_ssh_in_pods(namespace: str, context: Optional[str],
-                       new_nodes: List) -> None:
-    # Setting up ssh for the pod instance. This is already setup for
-    # the jump pod so it does not need to be run for it.
-    set_k8s_ssh_cmd = (
-        'set -ex; '
+    install_ssh_k8s_cmd = (
         'prefix_cmd() '
         '{ if [ $(id -u) -ne 0 ]; then echo "sudo"; else echo ""; fi; }; '
         'export DEBIAN_FRONTEND=noninteractive;'
-        '$(prefix_cmd) apt-get update;'
-        '$(prefix_cmd) apt install openssh-server rsync -y; '
+        'echo "Installing missing packages..."; '
+        'for i in {1..5}; do '
+        '  output=$($(prefix_cmd) apt install openssh-server rsync -y 2>&1); '
+        '  rc=$?; '
+        '  if [ $rc -eq 0 ]; then '
+        '    break; '
+        '  fi; '
+        '  echo "$output" | grep -qi "could not get lock" || '
+        '  grep -qi "Unable to acquire the dpkg frontend lock"; '
+        '  if [ $? -eq 0 ]; then '
+        '    echo "apt install failed due to lock, retrying. (Attempt $i/5)"; '
+        '    sleep 5; '
+        '  else '
+        '    echo "apt install failed for a non-lock reason: $output"; '
+        '    exit $rc; '
+        '  fi; '
+        'done; '
+        'if [ $rc -ne 0 ]; then '
+        '    echo "apt install failed after 5 attempts due to lock errors."; '
+        '    exit $rc; '
+        'fi; '
         '$(prefix_cmd) mkdir -p /var/run/sshd; '
         '$(prefix_cmd) '
         'sed -i "s/PermitRootLogin prohibit-password/PermitRootLogin yes/" '
@@ -456,24 +452,35 @@ def _setup_ssh_in_pods(namespace: str, context: Optional[str],
         # See https://www.educative.io/answers/error-mesg-ttyname-failed-inappropriate-ioctl-for-device  # pylint: disable=line-too-long
         '$(prefix_cmd) sed -i "s/mesg n/tty -s \\&\\& mesg n/" ~/.profile;')
 
-    def _setup_ssh_thread(new_node):
+    pre_init_cmd = ('set -ex; ' + check_k8s_user_sudo_cmd +
+                    set_k8s_env_var_cmd + check_apt_update_complete_cmd +
+                    install_ssh_k8s_cmd)
+
+    def _pre_init_thread(new_node):
         pod_name = new_node.metadata.name
+        logger.info(f'{"-"*20}Start: Pre-init in pod {pod_name!r} {"-"*20}')
         runner = command_runner.KubernetesCommandRunner(
             ((namespace, context), pod_name))
-        logger.info(f'{"-"*20}Start: Set up SSH in pod {pod_name!r} {"-"*20}')
 
-        def _run_ssh_setup():
-            rc, stdout, _ = runner.run(set_k8s_ssh_cmd,
-                                       require_outputs=True,
-                                       stream_logs=False)
-            _raise_command_running_error('setup ssh', set_k8s_ssh_cmd, pod_name,
-                                         rc, stdout)
+        # Run the combined pre-init command
+        rc, stdout, _ = runner.run(pre_init_cmd,
+                                   require_outputs=True,
+                                   stream_logs=False)
+        if rc == exceptions.INSUFFICIENT_PRIVILEGES_CODE:
+            raise config_lib.KubernetesError(
+                'Insufficient system privileges detected. '
+                'Ensure the default user has root access or '
+                '"sudo" is installed and the user is added to the sudoers '
+                'from the image.')
+
+        op_name = 'pre-init'
+        _raise_command_running_error(op_name, pre_init_cmd, pod_name, rc,
+                                     stdout)
 
-        _run_function_with_retries(_run_ssh_setup,
-                                   f'setup ssh in pod {pod_name!r}')
-        logger.info(f'{"-"*20}End: Set up SSH in pod {pod_name!r} {"-"*20}')
+        logger.info(f'{"-"*20}End: Pre-init in pod {pod_name!r} {"-"*20}')
 
-    subprocess_utils.run_in_parallel(_setup_ssh_thread, new_nodes, NUM_THREADS)
+    # Run pre_init in parallel across all new_nodes
+    subprocess_utils.run_in_parallel(_pre_init_thread, new_nodes, NUM_THREADS)
 
 
 def _label_pod(namespace: str, context: Optional[str], pod_name: str,
@@ -724,13 +731,8 @@ def _create_pods(region: str, cluster_name_on_cloud: str,
                      f'pods: {list(uninitialized_pods.keys())}')
         uninitialized_pods_list = list(uninitialized_pods.values())
 
-        # Setup SSH and environment variables in pods.
-        # Make sure commands used in these methods are generic and work
-        # on most base images. E.g., do not use Python, since that may not
-        # be installed by default.
-        _check_user_privilege(namespace, context, uninitialized_pods_list)
-        _setup_ssh_in_pods(namespace, context, uninitialized_pods_list)
-        _set_env_vars_in_pods(namespace, context, uninitialized_pods_list)
+        # Run pre-init steps in the pod.
+        pre_init(namespace, context, uninitialized_pods_list)
 
         for pod in uninitialized_pods.values():
             _label_pod(namespace,

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/provision/kubernetes/utils.py

@@ -1801,13 +1801,22 @@ def get_kubernetes_node_info(
     number of GPUs available on the node and the number of free GPUs on the
     node.
 
+    If the user does not have sufficient permissions to list pods in all
+    namespaces, the function will return free GPUs as -1.
+
     Returns:
         Dict[str, KubernetesNodeInfo]: Dictionary containing the node name as
             key and the KubernetesNodeInfo object as value
     """
     nodes = get_kubernetes_nodes(context)
     # Get the pods to get the real-time resource usage
-    pods = get_all_pods_in_kubernetes_cluster(context)
+    try:
+        pods = get_all_pods_in_kubernetes_cluster(context)
+    except kubernetes.api_exception() as e:
+        if e.status == 403:
+            pods = None
+        else:
+            raise
 
     label_formatter, _ = detect_gpu_label_formatter(context)
     if not label_formatter:
@@ -1828,19 +1837,22 @@ def get_kubernetes_node_info(
         accelerator_count = int(node.status.allocatable.get(
             'nvidia.com/gpu', 0))
 
-        for pod in pods:
-            # Get all the pods running on the node
-            if (pod.spec.node_name == node.metadata.name and
-                    pod.status.phase in ['Running', 'Pending']):
-                # Iterate over all the containers in the pod and sum the
-                # GPU requests
-                for container in pod.spec.containers:
-                    if container.resources.requests:
-                        allocated_qty += int(
-                            container.resources.requests.get(
-                                'nvidia.com/gpu', 0))
-
-        accelerators_available = accelerator_count - allocated_qty
+        if pods is None:
+            accelerators_available = -1
+
+        else:
+            for pod in pods:
+                # Get all the pods running on the node
+                if (pod.spec.node_name == node.metadata.name and
+                        pod.status.phase in ['Running', 'Pending']):
+                    # Iterate over all the containers in the pod and sum the
+                    # GPU requests
+                    for container in pod.spec.containers:
+                        if container.resources.requests:
+                            allocated_qty += int(
+                                container.resources.requests.get(
+                                    'nvidia.com/gpu', 0))
+            accelerators_available = accelerator_count - allocated_qty
 
         node_info_dict[node.metadata.name] = KubernetesNodeInfo(
             name=node.metadata.name,

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/serve/__init__.py

@@ -11,6 +11,7 @@ from sky.serve.core import tail_logs
 from sky.serve.core import terminate_replica
 from sky.serve.core import up
 from sky.serve.core import update
+from sky.serve.load_balancing_policies import LB_POLICIES
 from sky.serve.serve_state import ReplicaStatus
 from sky.serve.serve_state import ServiceStatus
 from sky.serve.serve_utils import DEFAULT_UPDATE_MODE
@@ -35,6 +36,7 @@ __all__ = [
     'get_endpoint',
     'INITIAL_VERSION',
     'LB_CONTROLLER_SYNC_INTERVAL_SECONDS',
+    'LB_POLICIES',
     'ReplicaStatus',
     'ServiceComponent',
     'ServiceStatus',

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/serve/load_balancer.py

@@ -2,7 +2,7 @@
 import asyncio
 import logging
 import threading
-from typing import Dict, Union
+from typing import Dict, Optional, Union
 
 import aiohttp
 import fastapi
@@ -27,18 +27,24 @@ class SkyServeLoadBalancer:
     policy.
     """
 
-    def __init__(self, controller_url: str, load_balancer_port: int) -> None:
+    def __init__(self,
+                 controller_url: str,
+                 load_balancer_port: int,
+                 load_balancing_policy_name: Optional[str] = None) -> None:
         """Initialize the load balancer.
 
         Args:
             controller_url: The URL of the controller.
             load_balancer_port: The port where the load balancer listens to.
+            load_balancing_policy_name: The name of the load balancing policy
+                to use. Defaults to None.
         """
         self._app = fastapi.FastAPI()
         self._controller_url: str = controller_url
         self._load_balancer_port: int = load_balancer_port
-        self._load_balancing_policy: lb_policies.LoadBalancingPolicy = (
-            lb_policies.RoundRobinPolicy())
+        # Use the registry to create the load balancing policy
+        self._load_balancing_policy = lb_policies.LoadBalancingPolicy.make(
+            load_balancing_policy_name)
         self._request_aggregator: serve_utils.RequestsAggregator = (
             serve_utils.RequestTimestamp())
         # TODO(tian): httpx.Client has a resource limit of 100 max connections
@@ -223,9 +229,21 @@ class SkyServeLoadBalancer:
         uvicorn.run(self._app, host='0.0.0.0', port=self._load_balancer_port)
 
 
-def run_load_balancer(controller_addr: str, load_balancer_port: int):
-    load_balancer = SkyServeLoadBalancer(controller_url=controller_addr,
-                                         load_balancer_port=load_balancer_port)
+def run_load_balancer(controller_addr: str,
+                      load_balancer_port: int,
+                      load_balancing_policy_name: Optional[str] = None) -> None:
+    """ Run the load balancer.
+
+    Args:
+        controller_addr: The address of the controller.
+        load_balancer_port: The port where the load balancer listens to.
+        policy_name: The name of the load balancing policy to use. Defaults to
+            None.
+    """
+    load_balancer = SkyServeLoadBalancer(
+        controller_url=controller_addr,
+        load_balancer_port=load_balancer_port,
+        load_balancing_policy_name=load_balancing_policy_name)
     load_balancer.run()
 
 
@@ -241,5 +259,13 @@ if __name__ == '__main__':
                         required=True,
                         default=8890,
                         help='The port where the load balancer listens to.')
+    available_policies = list(lb_policies.LB_POLICIES.keys())
+    parser.add_argument(
+        '--load-balancing-policy',
+        choices=available_policies,
+        default='round_robin',
+        help=f'The load balancing policy to use. Available policies: '
+        f'{", ".join(available_policies)}.')
     args = parser.parse_args()
-    run_load_balancer(args.controller_addr, args.load_balancer_port)
+    run_load_balancer(args.controller_addr, args.load_balancer_port,
+                      args.load_balancing_policy)

{skypilot_nightly-1.0.0.dev20241110 → skypilot_nightly-1.0.0.dev20241112}/sky/serve/load_balancing_policies.py

@@ -10,6 +10,10 @@ if typing.TYPE_CHECKING:
 
 logger = sky_logging.init_logger(__name__)
 
+# Define a registry for load balancing policies
+LB_POLICIES = {}
+DEFAULT_LB_POLICY = None
+
 
 def _request_repr(request: 'fastapi.Request') -> str:
     return ('<Request '
@@ -25,6 +29,24 @@ class LoadBalancingPolicy:
     def __init__(self) -> None:
         self.ready_replicas: List[str] = []
 
+    def __init_subclass__(cls, name: str, default: bool = False):
+        LB_POLICIES[name] = cls
+        if default:
+            global DEFAULT_LB_POLICY
+            assert DEFAULT_LB_POLICY is None, (
+                'Only one policy can be default.')
+            DEFAULT_LB_POLICY = name
+
+    @classmethod
+    def make(cls, policy_name: Optional[str] = None) -> 'LoadBalancingPolicy':
+        """Create a load balancing policy from a name."""
+        if policy_name is None:
+            policy_name = DEFAULT_LB_POLICY
+
+        if policy_name not in LB_POLICIES:
+            raise ValueError(f'Unknown load balancing policy: {policy_name}')
+        return LB_POLICIES[policy_name]()
+
     def set_ready_replicas(self, ready_replicas: List[str]) -> None:
         raise NotImplementedError
 
@@ -44,7 +66,7 @@ class LoadBalancingPolicy:
         raise NotImplementedError
 
 
-class RoundRobinPolicy(LoadBalancingPolicy):
+class RoundRobinPolicy(LoadBalancingPolicy, name='round_robin', default=True):
     """Round-robin load balancing policy."""
 
     def __init__(self) -> None: