ob-metaflow-stubs 6.0.10.16__py2.py3-none-any.whl → 6.0.10.18__py2.py3-none-any.whl

metaflow-stubs/__init__.pyi

@@ -1,7 +1,7 @@
 ######################################################################################################
 #                                 Auto-generated Metaflow stub file                                  #
-# MF version: 2.18.11.1+obcheckpoint(0.2.8);ob(v1)                                                   #
-# Generated on 2025-10-13T07:07:26.927215                                                            #
+# MF version: 2.18.12.1+obcheckpoint(0.2.8);ob(v1)                                                   #
+# Generated on 2025-10-20T19:13:33.388213                                                            #
 ######################################################################################################
 
 from __future__ import annotations
@@ -40,18 +40,18 @@ from .user_decorators.user_step_decorator import StepMutator as StepMutator
 from .user_decorators.user_step_decorator import user_step_decorator as user_step_decorator
 from .user_decorators.user_flow_decorator import FlowMutator as FlowMutator
 from . import cards as cards
-from . import tuple_util as tuple_util
 from . import metaflow_git as metaflow_git
+from . import tuple_util as tuple_util
 from . import events as events
 from . import runner as runner
 from . import plugins as plugins
 from .mf_extensions.outerbounds.toplevel.global_aliases_for_metaflow_package import S3 as S3
 from . import includefile as includefile
 from .includefile import IncludeFile as IncludeFile
-from .plugins.parsers import yaml_parser as yaml_parser
+from .plugins.pypi.parsers import pyproject_toml_parser as pyproject_toml_parser
 from .plugins.pypi.parsers import conda_environment_yml_parser as conda_environment_yml_parser
 from .plugins.pypi.parsers import requirements_txt_parser as requirements_txt_parser
-from .plugins.pypi.parsers import pyproject_toml_parser as pyproject_toml_parser
+from .plugins.parsers import yaml_parser as yaml_parser
 from . import client as client
 from .client.core import namespace as namespace
 from .client.core import get_namespace as get_namespace
@@ -169,101 +169,13 @@ def step(f: typing.Union[typing.Callable[[FlowSpecDerived], None], typing.Callab
     """
     ...
 
-def vllm(*, model: str, backend: str, openai_api_server: bool, debug: bool, card_refresh_interval: int, max_retries: int, retry_alert_frequency: int, engine_args: dict) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
-    """
-    This decorator is used to run vllm APIs as Metaflow task sidecars.
-    
-    User code call
-    --------------
-    @vllm(
-        model="...",
-        ...
-    )
-    
-    Valid backend options
-    ---------------------
-    - 'local': Run as a separate process on the local task machine.
-    
-    Valid model options
-    -------------------
-    Any HuggingFace model identifier, e.g. 'meta-llama/Llama-3.2-1B'
-    
-    NOTE: vLLM's OpenAI-compatible server serves ONE model per server instance.
-    If you need multiple models, you must create multiple @vllm decorators.
-    
-    
-    Parameters
-    ----------
-    model: str
-        HuggingFace model identifier to be served by vLLM.
-    backend: str
-        Determines where and how to run the vLLM process.
-    openai_api_server: bool
-        Whether to use OpenAI-compatible API server mode (subprocess) instead of native engine.
-        Default is False (uses native engine).
-        Set to True for backward compatibility with existing code.
-    debug: bool
-        Whether to turn on verbose debugging logs.
-    card_refresh_interval: int
-        Interval in seconds for refreshing the vLLM status card.
-        Only used when openai_api_server=True.
-    max_retries: int
-        Maximum number of retries checking for vLLM server startup.
-        Only used when openai_api_server=True.
-    retry_alert_frequency: int
-        Frequency of alert logs for vLLM server startup retries.
-        Only used when openai_api_server=True.
-    engine_args : dict
-        Additional keyword arguments to pass to the vLLM engine.
-        For example, `tensor_parallel_size=2`.
-    """
-    ...
-
-def ollama(*, models: list, backend: str, force_pull: bool, cache_update_policy: str, force_cache_update: bool, debug: bool, circuit_breaker_config: dict, timeout_config: dict) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def nebius_s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    This decorator is used to run Ollama APIs as Metaflow task sidecars.
-    
-    User code call
-    --------------
-    @ollama(
-        models=[...],
-        ...
-    )
-    
-    Valid backend options
-    ---------------------
-    - 'local': Run as a separate process on the local task machine.
-    - (TODO) 'managed': Outerbounds hosts and selects compute provider.
-    - (TODO) 'remote': Spin up separate instance to serve Ollama models.
-    
-    Valid model options
-    -------------------
-    Any model here https://ollama.com/search, e.g. 'llama3.2', 'llama3.3'
+    `@nebius_s3_proxy` is a Nebius-specific S3 Proxy decorator for routing S3 requests through a local proxy service.
+    It exists to make it easier for users to know that this decorator should only be used with
+    a Neo Cloud like Nebius. The underlying mechanics of the decorator is the same as the `@s3_proxy`:
     
     
-    Parameters
-    ----------
-    models: list[str]
-        List of Ollama containers running models in sidecars.
-    backend: str
-        Determines where and how to run the Ollama process.
-    force_pull: bool
-        Whether to run `ollama pull` no matter what, or first check the remote cache in Metaflow datastore for this model key.
-    cache_update_policy: str
-        Cache update policy: "auto", "force", or "never".
-    force_cache_update: bool
-        Simple override for "force" cache update policy.
-    debug: bool
-        Whether to turn on verbose debugging logs.
-    circuit_breaker_config: dict
-        Configuration for circuit breaker protection. Keys: failure_threshold, recovery_timeout, reset_timeout.
-    timeout_config: dict
-        Configuration for various operation timeouts. Keys: pull, stop, health_check, install, server_startup.
-    """
-    ...
-
-def s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
-    """
     Set up an S3 proxy that caches objects in an external, S3‑compatible bucket
     for S3 read and write requests.
     
@@ -321,236 +233,353 @@ def s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typin
     """
     ...
 
-def kubernetes(*, cpu: int = 1, memory: int = 4096, disk: int = 10240, image: typing.Optional[str] = None, image_pull_policy: str = 'KUBERNETES_IMAGE_PULL_POLICY', image_pull_secrets: typing.List[str] = [], service_account: str = 'METAFLOW_KUBERNETES_SERVICE_ACCOUNT', secrets: typing.Optional[typing.List[str]] = None, node_selector: typing.Union[typing.Dict[str, str], str, None] = None, namespace: str = 'METAFLOW_KUBERNETES_NAMESPACE', gpu: typing.Optional[int] = None, gpu_vendor: str = 'KUBERNETES_GPU_VENDOR', tolerations: typing.List[typing.Dict[str, str]] = [], labels: typing.Dict[str, str] = 'METAFLOW_KUBERNETES_LABELS', annotations: typing.Dict[str, str] = 'METAFLOW_KUBERNETES_ANNOTATIONS', use_tmpfs: bool = False, tmpfs_tempdir: bool = True, tmpfs_size: typing.Optional[int] = None, tmpfs_path: typing.Optional[str] = '/metaflow_temp', persistent_volume_claims: typing.Optional[typing.Dict[str, str]] = None, shared_memory: typing.Optional[int] = None, port: typing.Optional[int] = None, compute_pool: typing.Optional[str] = None, hostname_resolution_timeout: int = 600, qos: str = 'Burstable', security_context: typing.Optional[typing.Dict[str, typing.Any]] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+@typing.overload
+def conda(*, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies that this step should execute on Kubernetes.
+    Specifies the Conda environment for the step.
+    
+    Information in this decorator will augment any
+    attributes set in the `@conda_base` flow-level decorator. Hence,
+    you can use `@conda_base` to set packages required by all
+    steps and use `@conda` to specify step-specific overrides.
     
     
     Parameters
     ----------
-    cpu : int, default 1
-        Number of CPUs required for this step. If `@resources` is
-        also present, the maximum value from all decorators is used.
-    memory : int, default 4096
-        Memory size (in MB) required for this step. If
-        `@resources` is also present, the maximum value from all decorators is
-        used.
-    disk : int, default 10240
-        Disk size (in MB) required for this step. If
-        `@resources` is also present, the maximum value from all decorators is
-        used.
-    image : str, optional, default None
-        Docker image to use when launching on Kubernetes. If not specified, and
-        METAFLOW_KUBERNETES_CONTAINER_IMAGE is specified, that image is used. If
-        not, a default Docker image mapping to the current version of Python is used.
-    image_pull_policy: str, default KUBERNETES_IMAGE_PULL_POLICY
-        If given, the imagePullPolicy to be applied to the Docker image of the step.
-    image_pull_secrets: List[str], default []
-        The default is extracted from METAFLOW_KUBERNETES_IMAGE_PULL_SECRETS.
-        Kubernetes image pull secrets to use when pulling container images
-        in Kubernetes.
-    service_account : str, default METAFLOW_KUBERNETES_SERVICE_ACCOUNT
-        Kubernetes service account to use when launching pod in Kubernetes.
-    secrets : List[str], optional, default None
-        Kubernetes secrets to use when launching pod in Kubernetes. These
-        secrets are in addition to the ones defined in `METAFLOW_KUBERNETES_SECRETS`
-        in Metaflow configuration.
-    node_selector: Union[Dict[str,str], str], optional, default None
-        Kubernetes node selector(s) to apply to the pod running the task.
-        Can be passed in as a comma separated string of values e.g.
-        'kubernetes.io/os=linux,kubernetes.io/arch=amd64' or as a dictionary
-        {'kubernetes.io/os': 'linux', 'kubernetes.io/arch': 'amd64'}
-    namespace : str, default METAFLOW_KUBERNETES_NAMESPACE
-        Kubernetes namespace to use when launching pod in Kubernetes.
-    gpu : int, optional, default None
-        Number of GPUs required for this step. A value of zero implies that
-        the scheduled node should not have GPUs.
-    gpu_vendor : str, default KUBERNETES_GPU_VENDOR
-        The vendor of the GPUs to be used for this step.
-    tolerations : List[Dict[str,str]], default []
-        The default is extracted from METAFLOW_KUBERNETES_TOLERATIONS.
-        Kubernetes tolerations to use when launching pod in Kubernetes.
-    labels: Dict[str, str], default: METAFLOW_KUBERNETES_LABELS
-        Kubernetes labels to use when launching pod in Kubernetes.
-    annotations: Dict[str, str], default: METAFLOW_KUBERNETES_ANNOTATIONS
-        Kubernetes annotations to use when launching pod in Kubernetes.
-    use_tmpfs : bool, default False
-        This enables an explicit tmpfs mount for this step.
-    tmpfs_tempdir : bool, default True
-        sets METAFLOW_TEMPDIR to tmpfs_path if set for this step.
-    tmpfs_size : int, optional, default: None
-        The value for the size (in MiB) of the tmpfs mount for this step.
-        This parameter maps to the `--tmpfs` option in Docker. Defaults to 50% of the
-        memory allocated for this step.
-    tmpfs_path : str, optional, default /metaflow_temp
-        Path to tmpfs mount for this step.
-    persistent_volume_claims : Dict[str, str], optional, default None
-        A map (dictionary) of persistent volumes to be mounted to the pod for this step. The map is from persistent
-        volumes to the path to which the volume is to be mounted, e.g., `{'pvc-name': '/path/to/mount/on'}`.
-    shared_memory: int, optional
-        Shared memory size (in MiB) required for this step
-    port: int, optional
-        Port number to specify in the Kubernetes job object
-    compute_pool : str, optional, default None
-        Compute pool to be used for for this step.
-        If not specified, any accessible compute pool within the perimeter is used.
-    hostname_resolution_timeout: int, default 10 * 60
-        Timeout in seconds for the workers tasks in the gang scheduled cluster to resolve the hostname of control task.
-        Only applicable when @parallel is used.
-    qos: str, default: Burstable
-        Quality of Service class to assign to the pod. Supported values are: Guaranteed, Burstable, BestEffort
-    
-    security_context: Dict[str, Any], optional, default None
-        Container security context. Applies to the task container. Allows the following keys:
-        - privileged: bool, optional, default None
-        - allow_privilege_escalation: bool, optional, default None
-        - run_as_user: int, optional, default None
-        - run_as_group: int, optional, default None
-        - run_as_non_root: bool, optional, default None
+    packages : Dict[str, str], default {}
+        Packages to use for this step. The key is the name of the package
+        and the value is the version to use.
+    libraries : Dict[str, str], default {}
+        Supported for backward compatibility. When used with packages, packages will take precedence.
+    python : str, optional, default None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    disabled : bool, default False
+        If set to True, disables @conda.
     """
     ...
 
 @typing.overload
-def app_deploy(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    """
-    Decorator prototype for all step decorators. This function gets specialized
-    and imported for all decorators types by _import_plugin_decorators().
-    """
+def conda(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def app_deploy(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def conda(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def app_deploy(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
+def conda(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False):
     """
-    Decorator prototype for all step decorators. This function gets specialized
-    and imported for all decorators types by _import_plugin_decorators().
+    Specifies the Conda environment for the step.
+    
+    Information in this decorator will augment any
+    attributes set in the `@conda_base` flow-level decorator. Hence,
+    you can use `@conda_base` to set packages required by all
+    steps and use `@conda` to specify step-specific overrides.
+    
+    
+    Parameters
+    ----------
+    packages : Dict[str, str], default {}
+        Packages to use for this step. The key is the name of the package
+        and the value is the version to use.
+    libraries : Dict[str, str], default {}
+        Supported for backward compatibility. When used with packages, packages will take precedence.
+    python : str, optional, default None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    disabled : bool, default False
+        If set to True, disables @conda.
     """
     ...
 
-def huggingface_hub(*, temp_dir_root: typing.Optional[str] = None, cache_scope: typing.Optional[str] = None, load: typing.Union[typing.List[str], typing.List[typing.Tuple[typing.Dict, str]], typing.List[typing.Tuple[str, str]], typing.List[typing.Dict], None]) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+@typing.overload
+def checkpoint(*, load_policy: str = 'fresh', temp_dir_root: str = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Decorator that helps cache, version, and store models/datasets from the Hugging Face Hub.
+    Enables checkpointing for a step.
     
-    Examples
-    --------
+    > Examples
+    
+    - Saving Checkpoints
     
     ```python
-    # **Usage: creating references to models from the Hugging Face Hub that may be loaded in downstream steps**
-    @huggingface_hub
+    @checkpoint
     @step
-    def pull_model_from_huggingface(self):
-        # `current.huggingface_hub.snapshot_download` downloads the model from the Hugging Face Hub
-        # and saves it in the backend storage based on the model's `repo_id`. If there exists a model
-        # with the same `repo_id` in the backend storage, it will not download the model again. The return
-        # value of the function is a reference to the model in the backend storage.
-        # This reference can be used to load the model in the subsequent steps via `@model(load=["llama_model"])`
-    
-        self.model_id = "mistralai/Mistral-7B-Instruct-v0.1"
-        self.llama_model = current.huggingface_hub.snapshot_download(
-            repo_id=self.model_id,
-            allow_patterns=["*.safetensors", "*.json", "tokenizer.*"],
-        )
-        self.next(self.train)
+    def train(self):
+        model = create_model(self.parameters, checkpoint_path = None)
+        for i in range(self.epochs):
+            # some training logic
+            loss = model.train(self.dataset)
+            if i % 10 == 0:
+                model.save(
+                    current.checkpoint.directory,
+                )
+                # saves the contents of the `current.checkpoint.directory` as a checkpoint
+                # and returns a reference dictionary to the checkpoint saved in the datastore
+                self.latest_checkpoint = current.checkpoint.save(
+                    name="epoch_checkpoint",
+                    metadata={
+                        "epoch": i,
+                        "loss": loss,
+                    }
+                )
+    ```
     
-    # **Usage: explicitly loading models at runtime from the Hugging Face Hub or from cache (from Metaflow's datastore)**
-    @huggingface_hub
+    - Using Loaded Checkpoints
+    
+    ```python
+    @retry(times=3)
+    @checkpoint
     @step
-    def run_training(self):
-        # Temporary directory (auto-cleaned on exit)
-        with current.huggingface_hub.load(
-            repo_id="google-bert/bert-base-uncased",
-            allow_patterns=["*.bin"],
-        ) as local_path:
-            # Use files under local_path
-            train_model(local_path)
+    def train(self):
+        # Assume that the task has restarted and the previous attempt of the task
+        # saved a checkpoint
+        checkpoint_path = None
+        if current.checkpoint.is_loaded: # Check if a checkpoint is loaded
+            print("Loaded checkpoint from the previous attempt")
+            checkpoint_path = current.checkpoint.directory
+    
+        model = create_model(self.parameters, checkpoint_path = checkpoint_path)
+        for i in range(self.epochs):
             ...
+    ```
     
-    # **Usage: loading models directly from the Hugging Face Hub or from cache (from Metaflow's datastore)**
     
-    @huggingface_hub(load=["mistralai/Mistral-7B-Instruct-v0.1"])
-    @step
-    def pull_model_from_huggingface(self):
-        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
+    Parameters
+    ----------
+    load_policy : str, default: "fresh"
+        The policy for loading the checkpoint. The following policies are supported:
+            - "eager": Loads the the latest available checkpoint within the namespace.
+            With this mode, the latest checkpoint written by any previous task (can be even a different run) of the step
+            will be loaded at the start of the task.
+            - "none": Do not load any checkpoint
+            - "fresh": Loads the lastest checkpoint created within the running Task.
+            This mode helps loading checkpoints across various retry attempts of the same task.
+            With this mode, no checkpoint will be loaded at the start of a task but any checkpoints
+            created within the task will be loaded when the task is retries execution on failure.
     
-    @huggingface_hub(load=[("mistralai/Mistral-7B-Instruct-v0.1", "/my-directory"), ("myorg/mistral-lora", "/my-lora-directory")])
+    temp_dir_root : str, default: None
+        The root directory under which `current.checkpoint.directory` will be created.
+    """
+    ...
+
+@typing.overload
+def checkpoint(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+    ...
+
+@typing.overload
+def checkpoint(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+    ...
+
+def checkpoint(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, load_policy: str = 'fresh', temp_dir_root: str = None):
+    """
+    Enables checkpointing for a step.
+    
+    > Examples
+    
+    - Saving Checkpoints
+    
+    ```python
+    @checkpoint
     @step
-    def finetune_model(self):
-        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
-        # path_to_model will be /my-directory
+    def train(self):
+        model = create_model(self.parameters, checkpoint_path = None)
+        for i in range(self.epochs):
+            # some training logic
+            loss = model.train(self.dataset)
+            if i % 10 == 0:
+                model.save(
+                    current.checkpoint.directory,
+                )
+                # saves the contents of the `current.checkpoint.directory` as a checkpoint
+                # and returns a reference dictionary to the checkpoint saved in the datastore
+                self.latest_checkpoint = current.checkpoint.save(
+                    name="epoch_checkpoint",
+                    metadata={
+                        "epoch": i,
+                        "loss": loss,
+                    }
+                )
+    ```
     
+    - Using Loaded Checkpoints
     
-    # Takes all the arguments passed to `snapshot_download`
-    # except for `local_dir`
-    @huggingface_hub(load=[
-        {
-            "repo_id": "mistralai/Mistral-7B-Instruct-v0.1",
-        },
-        {
-            "repo_id": "myorg/mistral-lora",
-            "repo_type": "model",
-        },
-    ])
+    ```python
+    @retry(times=3)
+    @checkpoint
     @step
-    def finetune_model(self):
-        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
-        # path_to_model will be /my-directory
+    def train(self):
+        # Assume that the task has restarted and the previous attempt of the task
+        # saved a checkpoint
+        checkpoint_path = None
+        if current.checkpoint.is_loaded: # Check if a checkpoint is loaded
+            print("Loaded checkpoint from the previous attempt")
+            checkpoint_path = current.checkpoint.directory
+    
+        model = create_model(self.parameters, checkpoint_path = checkpoint_path)
+        for i in range(self.epochs):
+            ...
     ```
     
     
     Parameters
     ----------
-    temp_dir_root : str, optional
-        The root directory that will hold the temporary directory where objects will be downloaded.
+    load_policy : str, default: "fresh"
+        The policy for loading the checkpoint. The following policies are supported:
+            - "eager": Loads the the latest available checkpoint within the namespace.
+            With this mode, the latest checkpoint written by any previous task (can be even a different run) of the step
+            will be loaded at the start of the task.
+            - "none": Do not load any checkpoint
+            - "fresh": Loads the lastest checkpoint created within the running Task.
+            This mode helps loading checkpoints across various retry attempts of the same task.
+            With this mode, no checkpoint will be loaded at the start of a task but any checkpoints
+            created within the task will be loaded when the task is retries execution on failure.
     
-    cache_scope : str, optional
-        The scope of the cache. Can be `checkpoint` / `flow` / `global`.
-            - `checkpoint` (default): All repos are stored like objects saved by `@checkpoint`.
-                i.e., the cached path is derived from the namespace, flow, step, and Metaflow foreach iteration.
-                Any repo downloaded under this scope will only be retrieved from the cache when the step runs under the same namespace in the same flow (at the same foreach index).
+    temp_dir_root : str, default: None
+        The root directory under which `current.checkpoint.directory` will be created.
+    """
+    ...
+
+@typing.overload
+def card(*, type: str = 'default', id: typing.Optional[str] = None, options: typing.Dict[str, typing.Any] = {}, timeout: int = 45) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+    """
+    Creates a human-readable report, a Metaflow Card, after this step completes.
     
-            - `flow`: All repos are cached under the flow, regardless of namespace.
-                i.e., the cached path is derived solely from the flow name.
-                When to use this mode: (1) Multiple users are executing the same flow and want shared access to the repos cached by the decorator. (2) Multiple versions of a flow are deployed, all needing access to the same repos cached by the decorator.
+    Note that you may add multiple `@card` decorators in a step with different parameters.
     
-            - `global`: All repos are cached under a globally static path.
-                i.e., the base path of the cache is static and all repos are stored under it.
-                When to use this mode:
-                    - All repos from the Hugging Face Hub need to be shared by users across all flow executions.
-            - Each caching scope comes with its own trade-offs:
-                - `checkpoint`:
-                    - Has explicit control over when caches are populated (controlled by the same flow that has the `@huggingface_hub` decorator) but ends up hitting the Hugging Face Hub more often if there are many users/namespaces/steps.
-                    - Since objects are written on a `namespace/flow/step` basis, the blast radius of a bad checkpoint is limited to a particular flow in a namespace.
-                - `flow`:
-                    - Has less control over when caches are populated (can be written by any execution instance of a flow from any namespace) but results in more cache hits.
-                    - The blast radius of a bad checkpoint is limited to all runs of a particular flow.
-                    - It doesn't promote cache reuse across flows.
-                - `global`:
-                    - Has no control over when caches are populated (can be written by any flow execution) but has the highest cache hit rate.
-                    - It promotes cache reuse across flows.
-                    - The blast radius of a bad checkpoint spans every flow that could be using a particular repo.
     
-    load: Union[List[str], List[Tuple[Dict, str]], List[Tuple[str, str]], List[Dict], None]
-        The list of repos (models/datasets) to load.
+    Parameters
+    ----------
+    type : str, default 'default'
+        Card type.
+    id : str, optional, default None
+        If multiple cards are present, use this id to identify this card.
+    options : Dict[str, Any], default {}
+        Options passed to the card. The contents depend on the card type.
+    timeout : int, default 45
+        Interrupt reporting if it takes more than this many seconds.
+    """
+    ...
+
+@typing.overload
+def card(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+    ...
+
+@typing.overload
+def card(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+    ...
+
+def card(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, type: str = 'default', id: typing.Optional[str] = None, options: typing.Dict[str, typing.Any] = {}, timeout: int = 45):
+    """
+    Creates a human-readable report, a Metaflow Card, after this step completes.
     
-        Loaded repos can be accessed via `current.huggingface_hub.loaded`. If load is set, then the following happens:
+    Note that you may add multiple `@card` decorators in a step with different parameters.
     
-        - If repo (model/dataset) is not found in the datastore:
-            - Downloads the repo from Hugging Face Hub to a temporary directory (or uses specified path) for local access
-            - Stores it in Metaflow's datastore (s3/gcs/azure etc.) with a unique name based on repo_type/repo_id
-                - All HF models loaded for a `@step` will be cached separately under flow/step/namespace.
     
-        - If repo is found in the datastore:
-            - Loads it directly from datastore to local path (can be temporary directory or specified path)
+    Parameters
+    ----------
+    type : str, default 'default'
+        Card type.
+    id : str, optional, default None
+        If multiple cards are present, use this id to identify this card.
+    options : Dict[str, Any], default {}
+        Options passed to the card. The contents depend on the card type.
+    timeout : int, default 45
+        Interrupt reporting if it takes more than this many seconds.
     """
     ...
 
-def nebius_s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+@typing.overload
+def fast_bakery_internal(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     """
-    `@nebius_s3_proxy` is a Nebius-specific S3 Proxy decorator for routing S3 requests through a local proxy service.
+    Internal decorator to support Fast bakery
+    """
+    ...
+
+@typing.overload
+def fast_bakery_internal(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+    ...
+
+def fast_bakery_internal(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
+    """
+    Internal decorator to support Fast bakery
+    """
+    ...
+
+@typing.overload
+def app_deploy(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+    """
+    Decorator prototype for all step decorators. This function gets specialized
+    and imported for all decorators types by _import_plugin_decorators().
+    """
+    ...
+
+@typing.overload
+def app_deploy(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+    ...
+
+def app_deploy(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
+    """
+    Decorator prototype for all step decorators. This function gets specialized
+    and imported for all decorators types by _import_plugin_decorators().
+    """
+    ...
+
+@typing.overload
+def pypi(*, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+    """
+    Specifies the PyPI packages for the step.
+    
+    Information in this decorator will augment any
+    attributes set in the `@pyi_base` flow-level decorator. Hence,
+    you can use `@pypi_base` to set packages required by all
+    steps and use `@pypi` to specify step-specific overrides.
+    
+    
+    Parameters
+    ----------
+    packages : Dict[str, str], default: {}
+        Packages to use for this step. The key is the name of the package
+        and the value is the version to use.
+    python : str, optional, default: None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    """
+    ...
+
+@typing.overload
+def pypi(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+    ...
+
+@typing.overload
+def pypi(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+    ...
+
+def pypi(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None):
+    """
+    Specifies the PyPI packages for the step.
+    
+    Information in this decorator will augment any
+    attributes set in the `@pyi_base` flow-level decorator. Hence,
+    you can use `@pypi_base` to set packages required by all
+    steps and use `@pypi` to specify step-specific overrides.
+    
+    
+    Parameters
+    ----------
+    packages : Dict[str, str], default: {}
+        Packages to use for this step. The key is the name of the package
+        and the value is the version to use.
+    python : str, optional, default: None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    """
+    ...
+
+def coreweave_s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+    """
+    `@coreweave_s3_proxy` is a CoreWeave-specific S3 Proxy decorator for routing S3 requests through a local proxy service.
     It exists to make it easier for users to know that this decorator should only be used with
-    a Neo Cloud like Nebius. The underlying mechanics of the decorator is the same as the `@s3_proxy`:
+    a Neo Cloud like CoreWeave. The underlying mechanics of the decorator is the same as the `@s3_proxy`:
     
     
     Set up an S3 proxy that caches objects in an external, S3‑compatible bucket
@@ -611,162 +640,72 @@ def nebius_s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode
     ...
 
 @typing.overload
-def fast_bakery_internal(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def test_append_card(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     """
-    Internal decorator to support Fast bakery
+    A simple decorator that demonstrates using CardDecoratorInjector
+    to inject a card and render simple markdown content.
     """
     ...
 
 @typing.overload
-def fast_bakery_internal(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def fast_bakery_internal(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
-    """
-    Internal decorator to support Fast bakery
-    """
+def test_append_card(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def nvct(*, gpu: int, gpu_type: str) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def test_append_card(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
     """
-    Specifies that this step should execute on DGX cloud.
-    
-    
-    Parameters
-    ----------
-    gpu : int
-        Number of GPUs to use.
-    gpu_type : str
-        Type of Nvidia GPU to use.
+    A simple decorator that demonstrates using CardDecoratorInjector
+    to inject a card and render simple markdown content.
     """
     ...
 
 @typing.overload
-def model(*, load: typing.Union[typing.List[str], str, typing.List[typing.Tuple[str, typing.Optional[str]]]] = None, temp_dir_root: str = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def catch(*, var: typing.Optional[str] = None, print_exception: bool = True) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Enables loading / saving of models within a step.
-    
-    > Examples
-    - Saving Models
-    ```python
-    @model
-    @step
-    def train(self):
-        # current.model.save returns a dictionary reference to the model saved
-        self.my_model = current.model.save(
-            path_to_my_model,
-            label="my_model",
-            metadata={
-                "epochs": 10,
-                "batch-size": 32,
-                "learning-rate": 0.001,
-            }
-        )
-        self.next(self.test)
-    
-    @model(load="my_model")
-    @step
-    def test(self):
-        # `current.model.loaded` returns a dictionary of the loaded models
-        # where the key is the name of the artifact and the value is the path to the model
-        print(os.listdir(current.model.loaded["my_model"]))
-        self.next(self.end)
-    ```
+    Specifies that the step will success under all circumstances.
     
-    - Loading models
-    ```python
-    @step
-    def train(self):
-        # current.model.load returns the path to the model loaded
-        checkpoint_path = current.model.load(
-            self.checkpoint_key,
-        )
-        model_path = current.model.load(
-            self.model,
-        )
-        self.next(self.test)
-    ```
+    The decorator will create an optional artifact, specified by `var`, which
+    contains the exception raised. You can use it to detect the presence
+    of errors, indicating that all happy-path artifacts produced by the step
+    are missing.
     
     
     Parameters
     ----------
-    load : Union[List[str],str,List[Tuple[str,Union[str,None]]]], default: None
-        Artifact name/s referencing the models/checkpoints to load. Artifact names refer to the names of the instance variables set to `self`.
-        These artifact names give to `load` be reference objects or reference `key` string's from objects created by `current.checkpoint` / `current.model` / `current.huggingface_hub`.
-        If a list of tuples is provided, the first element is the artifact name and the second element is the path the artifact needs be unpacked on
-        the local filesystem. If the second element is None, the artifact will be unpacked in the current working directory.
-        If a string is provided, then the artifact corresponding to that name will be loaded in the current working directory.
-    
-    temp_dir_root : str, default: None
-        The root directory under which `current.model.loaded` will store loaded models
+    var : str, optional, default None
+        Name of the artifact in which to store the caught exception.
+        If not specified, the exception is not stored.
+    print_exception : bool, default True
+        Determines whether or not the exception is printed to
+        stdout when caught.
     """
     ...
 
 @typing.overload
-def model(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def catch(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def model(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def catch(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def model(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, load: typing.Union[typing.List[str], str, typing.List[typing.Tuple[str, typing.Optional[str]]]] = None, temp_dir_root: str = None):
+def catch(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, var: typing.Optional[str] = None, print_exception: bool = True):
     """
-    Enables loading / saving of models within a step.
-    
-    > Examples
-    - Saving Models
-    ```python
-    @model
-    @step
-    def train(self):
-        # current.model.save returns a dictionary reference to the model saved
-        self.my_model = current.model.save(
-            path_to_my_model,
-            label="my_model",
-            metadata={
-                "epochs": 10,
-                "batch-size": 32,
-                "learning-rate": 0.001,
-            }
-        )
-        self.next(self.test)
-    
-    @model(load="my_model")
-    @step
-    def test(self):
-        # `current.model.loaded` returns a dictionary of the loaded models
-        # where the key is the name of the artifact and the value is the path to the model
-        print(os.listdir(current.model.loaded["my_model"]))
-        self.next(self.end)
-    ```
+    Specifies that the step will success under all circumstances.
     
-    - Loading models
-    ```python
-    @step
-    def train(self):
-        # current.model.load returns the path to the model loaded
-        checkpoint_path = current.model.load(
-            self.checkpoint_key,
-        )
-        model_path = current.model.load(
-            self.model,
-        )
-        self.next(self.test)
-    ```
+    The decorator will create an optional artifact, specified by `var`, which
+    contains the exception raised. You can use it to detect the presence
+    of errors, indicating that all happy-path artifacts produced by the step
+    are missing.
     
     
     Parameters
     ----------
-    load : Union[List[str],str,List[Tuple[str,Union[str,None]]]], default: None
-        Artifact name/s referencing the models/checkpoints to load. Artifact names refer to the names of the instance variables set to `self`.
-        These artifact names give to `load` be reference objects or reference `key` string's from objects created by `current.checkpoint` / `current.model` / `current.huggingface_hub`.
-        If a list of tuples is provided, the first element is the artifact name and the second element is the path the artifact needs be unpacked on
-        the local filesystem. If the second element is None, the artifact will be unpacked in the current working directory.
-        If a string is provided, then the artifact corresponding to that name will be loaded in the current working directory.
-    
-    temp_dir_root : str, default: None
-        The root directory under which `current.model.loaded` will store loaded models
+    var : str, optional, default None
+        Name of the artifact in which to store the caught exception.
+        If not specified, the exception is not stored.
+    print_exception : bool, default True
+        Determines whether or not the exception is printed to
+        stdout when caught.
     """
     ...
 
@@ -826,498 +765,409 @@ def retry(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], ty
     ...
 
 @typing.overload
-def card(*, type: str = 'default', id: typing.Optional[str] = None, options: typing.Dict[str, typing.Any] = {}, timeout: int = 45) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
-    """
-    Creates a human-readable report, a Metaflow Card, after this step completes.
-    
-    Note that you may add multiple `@card` decorators in a step with different parameters.
-    
-    
-    Parameters
-    ----------
-    type : str, default 'default'
-        Card type.
-    id : str, optional, default None
-        If multiple cards are present, use this id to identify this card.
-    options : Dict[str, Any], default {}
-        Options passed to the card. The contents depend on the card type.
-    timeout : int, default 45
-        Interrupt reporting if it takes more than this many seconds.
-    """
-    ...
-
-@typing.overload
-def card(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    ...
-
-@typing.overload
-def card(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def card(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, type: str = 'default', id: typing.Optional[str] = None, options: typing.Dict[str, typing.Any] = {}, timeout: int = 45):
-    """
-    Creates a human-readable report, a Metaflow Card, after this step completes.
-    
-    Note that you may add multiple `@card` decorators in a step with different parameters.
-    
-    
-    Parameters
-    ----------
-    type : str, default 'default'
-        Card type.
-    id : str, optional, default None
-        If multiple cards are present, use this id to identify this card.
-    options : Dict[str, Any], default {}
-        Options passed to the card. The contents depend on the card type.
-    timeout : int, default 45
-        Interrupt reporting if it takes more than this many seconds.
-    """
-    ...
-
-@typing.overload
-def parallel(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    """
-    Decorator prototype for all step decorators. This function gets specialized
-    and imported for all decorators types by _import_plugin_decorators().
-    """
-    ...
-
-@typing.overload
-def parallel(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def parallel(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
-    """
-    Decorator prototype for all step decorators. This function gets specialized
-    and imported for all decorators types by _import_plugin_decorators().
-    """
-    ...
-
-@typing.overload
-def checkpoint(*, load_policy: str = 'fresh', temp_dir_root: str = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def model(*, load: typing.Union[typing.List[str], str, typing.List[typing.Tuple[str, typing.Optional[str]]]] = None, temp_dir_root: str = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Enables checkpointing for a step.
+    Enables loading / saving of models within a step.
     
     > Examples
-    
-    - Saving Checkpoints
-    
+    - Saving Models
     ```python
-    @checkpoint
+    @model
     @step
     def train(self):
-        model = create_model(self.parameters, checkpoint_path = None)
-        for i in range(self.epochs):
-            # some training logic
-            loss = model.train(self.dataset)
-            if i % 10 == 0:
-                model.save(
-                    current.checkpoint.directory,
-                )
-                # saves the contents of the `current.checkpoint.directory` as a checkpoint
-                # and returns a reference dictionary to the checkpoint saved in the datastore
-                self.latest_checkpoint = current.checkpoint.save(
-                    name="epoch_checkpoint",
-                    metadata={
-                        "epoch": i,
-                        "loss": loss,
-                    }
-                )
-    ```
+        # current.model.save returns a dictionary reference to the model saved
+        self.my_model = current.model.save(
+            path_to_my_model,
+            label="my_model",
+            metadata={
+                "epochs": 10,
+                "batch-size": 32,
+                "learning-rate": 0.001,
+            }
+        )
+        self.next(self.test)
     
-    - Using Loaded Checkpoints
+    @model(load="my_model")
+    @step
+    def test(self):
+        # `current.model.loaded` returns a dictionary of the loaded models
+        # where the key is the name of the artifact and the value is the path to the model
+        print(os.listdir(current.model.loaded["my_model"]))
+        self.next(self.end)
+    ```
     
+    - Loading models
     ```python
-    @retry(times=3)
-    @checkpoint
     @step
     def train(self):
-        # Assume that the task has restarted and the previous attempt of the task
-        # saved a checkpoint
-        checkpoint_path = None
-        if current.checkpoint.is_loaded: # Check if a checkpoint is loaded
-            print("Loaded checkpoint from the previous attempt")
-            checkpoint_path = current.checkpoint.directory
-    
-        model = create_model(self.parameters, checkpoint_path = checkpoint_path)
-        for i in range(self.epochs):
-            ...
+        # current.model.load returns the path to the model loaded
+        checkpoint_path = current.model.load(
+            self.checkpoint_key,
+        )
+        model_path = current.model.load(
+            self.model,
+        )
+        self.next(self.test)
     ```
     
     
     Parameters
     ----------
-    load_policy : str, default: "fresh"
-        The policy for loading the checkpoint. The following policies are supported:
-            - "eager": Loads the the latest available checkpoint within the namespace.
-            With this mode, the latest checkpoint written by any previous task (can be even a different run) of the step
-            will be loaded at the start of the task.
-            - "none": Do not load any checkpoint
-            - "fresh": Loads the lastest checkpoint created within the running Task.
-            This mode helps loading checkpoints across various retry attempts of the same task.
-            With this mode, no checkpoint will be loaded at the start of a task but any checkpoints
-            created within the task will be loaded when the task is retries execution on failure.
+    load : Union[List[str],str,List[Tuple[str,Union[str,None]]]], default: None
+        Artifact name/s referencing the models/checkpoints to load. Artifact names refer to the names of the instance variables set to `self`.
+        These artifact names give to `load` be reference objects or reference `key` string's from objects created by `current.checkpoint` / `current.model` / `current.huggingface_hub`.
+        If a list of tuples is provided, the first element is the artifact name and the second element is the path the artifact needs be unpacked on
+        the local filesystem. If the second element is None, the artifact will be unpacked in the current working directory.
+        If a string is provided, then the artifact corresponding to that name will be loaded in the current working directory.
     
     temp_dir_root : str, default: None
-        The root directory under which `current.checkpoint.directory` will be created.
+        The root directory under which `current.model.loaded` will store loaded models
     """
     ...
 
 @typing.overload
-def checkpoint(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def model(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def checkpoint(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def model(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def checkpoint(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, load_policy: str = 'fresh', temp_dir_root: str = None):
+def model(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, load: typing.Union[typing.List[str], str, typing.List[typing.Tuple[str, typing.Optional[str]]]] = None, temp_dir_root: str = None):
     """
-    Enables checkpointing for a step.
+    Enables loading / saving of models within a step.
     
     > Examples
-    
-    - Saving Checkpoints
-    
+    - Saving Models
     ```python
-    @checkpoint
+    @model
     @step
     def train(self):
-        model = create_model(self.parameters, checkpoint_path = None)
-        for i in range(self.epochs):
-            # some training logic
-            loss = model.train(self.dataset)
-            if i % 10 == 0:
-                model.save(
-                    current.checkpoint.directory,
-                )
-                # saves the contents of the `current.checkpoint.directory` as a checkpoint
-                # and returns a reference dictionary to the checkpoint saved in the datastore
-                self.latest_checkpoint = current.checkpoint.save(
-                    name="epoch_checkpoint",
-                    metadata={
-                        "epoch": i,
-                        "loss": loss,
-                    }
-                )
-    ```
+        # current.model.save returns a dictionary reference to the model saved
+        self.my_model = current.model.save(
+            path_to_my_model,
+            label="my_model",
+            metadata={
+                "epochs": 10,
+                "batch-size": 32,
+                "learning-rate": 0.001,
+            }
+        )
+        self.next(self.test)
     
-    - Using Loaded Checkpoints
+    @model(load="my_model")
+    @step
+    def test(self):
+        # `current.model.loaded` returns a dictionary of the loaded models
+        # where the key is the name of the artifact and the value is the path to the model
+        print(os.listdir(current.model.loaded["my_model"]))
+        self.next(self.end)
+    ```
     
+    - Loading models
     ```python
-    @retry(times=3)
-    @checkpoint
     @step
     def train(self):
-        # Assume that the task has restarted and the previous attempt of the task
-        # saved a checkpoint
-        checkpoint_path = None
-        if current.checkpoint.is_loaded: # Check if a checkpoint is loaded
-            print("Loaded checkpoint from the previous attempt")
-            checkpoint_path = current.checkpoint.directory
-    
-        model = create_model(self.parameters, checkpoint_path = checkpoint_path)
-        for i in range(self.epochs):
-            ...
+        # current.model.load returns the path to the model loaded
+        checkpoint_path = current.model.load(
+            self.checkpoint_key,
+        )
+        model_path = current.model.load(
+            self.model,
+        )
+        self.next(self.test)
     ```
     
     
     Parameters
     ----------
-    load_policy : str, default: "fresh"
-        The policy for loading the checkpoint. The following policies are supported:
-            - "eager": Loads the the latest available checkpoint within the namespace.
-            With this mode, the latest checkpoint written by any previous task (can be even a different run) of the step
-            will be loaded at the start of the task.
-            - "none": Do not load any checkpoint
-            - "fresh": Loads the lastest checkpoint created within the running Task.
-            This mode helps loading checkpoints across various retry attempts of the same task.
-            With this mode, no checkpoint will be loaded at the start of a task but any checkpoints
-            created within the task will be loaded when the task is retries execution on failure.
+    load : Union[List[str],str,List[Tuple[str,Union[str,None]]]], default: None
+        Artifact name/s referencing the models/checkpoints to load. Artifact names refer to the names of the instance variables set to `self`.
+        These artifact names give to `load` be reference objects or reference `key` string's from objects created by `current.checkpoint` / `current.model` / `current.huggingface_hub`.
+        If a list of tuples is provided, the first element is the artifact name and the second element is the path the artifact needs be unpacked on
+        the local filesystem. If the second element is None, the artifact will be unpacked in the current working directory.
+        If a string is provided, then the artifact corresponding to that name will be loaded in the current working directory.
     
     temp_dir_root : str, default: None
-        The root directory under which `current.checkpoint.directory` will be created.
+        The root directory under which `current.model.loaded` will store loaded models
     """
     ...
 
-@typing.overload
-def timeout(*, seconds: int = 0, minutes: int = 0, hours: int = 0) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def kubernetes(*, cpu: int = 1, memory: int = 4096, disk: int = 10240, image: typing.Optional[str] = None, image_pull_policy: str = 'KUBERNETES_IMAGE_PULL_POLICY', image_pull_secrets: typing.List[str] = [], service_account: str = 'METAFLOW_KUBERNETES_SERVICE_ACCOUNT', secrets: typing.Optional[typing.List[str]] = None, node_selector: typing.Union[typing.Dict[str, str], str, None] = None, namespace: str = 'METAFLOW_KUBERNETES_NAMESPACE', gpu: typing.Optional[int] = None, gpu_vendor: str = 'KUBERNETES_GPU_VENDOR', tolerations: typing.List[typing.Dict[str, str]] = [], labels: typing.Dict[str, str] = 'METAFLOW_KUBERNETES_LABELS', annotations: typing.Dict[str, str] = 'METAFLOW_KUBERNETES_ANNOTATIONS', use_tmpfs: bool = False, tmpfs_tempdir: bool = True, tmpfs_size: typing.Optional[int] = None, tmpfs_path: typing.Optional[str] = '/metaflow_temp', persistent_volume_claims: typing.Optional[typing.Dict[str, str]] = None, shared_memory: typing.Optional[int] = None, port: typing.Optional[int] = None, compute_pool: typing.Optional[str] = None, hostname_resolution_timeout: int = 600, qos: str = 'Burstable', security_context: typing.Optional[typing.Dict[str, typing.Any]] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies a timeout for your step.
-    
-    This decorator is useful if this step may hang indefinitely.
-    
-    This can be used in conjunction with the `@retry` decorator as well as the `@catch` decorator.
-    A timeout is considered to be an exception thrown by the step. It will cause the step to be
-    retried if needed and the exception will be caught by the `@catch` decorator, if present.
-    
-    Note that all the values specified in parameters are added together so if you specify
-    60 seconds and 1 hour, the decorator will have an effective timeout of 1 hour and 1 minute.
+    Specifies that this step should execute on Kubernetes.
     
     
     Parameters
     ----------
-    seconds : int, default 0
-        Number of seconds to wait prior to timing out.
-    minutes : int, default 0
-        Number of minutes to wait prior to timing out.
-    hours : int, default 0
-        Number of hours to wait prior to timing out.
+    cpu : int, default 1
+        Number of CPUs required for this step. If `@resources` is
+        also present, the maximum value from all decorators is used.
+    memory : int, default 4096
+        Memory size (in MB) required for this step. If
+        `@resources` is also present, the maximum value from all decorators is
+        used.
+    disk : int, default 10240
+        Disk size (in MB) required for this step. If
+        `@resources` is also present, the maximum value from all decorators is
+        used.
+    image : str, optional, default None
+        Docker image to use when launching on Kubernetes. If not specified, and
+        METAFLOW_KUBERNETES_CONTAINER_IMAGE is specified, that image is used. If
+        not, a default Docker image mapping to the current version of Python is used.
+    image_pull_policy: str, default KUBERNETES_IMAGE_PULL_POLICY
+        If given, the imagePullPolicy to be applied to the Docker image of the step.
+    image_pull_secrets: List[str], default []
+        The default is extracted from METAFLOW_KUBERNETES_IMAGE_PULL_SECRETS.
+        Kubernetes image pull secrets to use when pulling container images
+        in Kubernetes.
+    service_account : str, default METAFLOW_KUBERNETES_SERVICE_ACCOUNT
+        Kubernetes service account to use when launching pod in Kubernetes.
+    secrets : List[str], optional, default None
+        Kubernetes secrets to use when launching pod in Kubernetes. These
+        secrets are in addition to the ones defined in `METAFLOW_KUBERNETES_SECRETS`
+        in Metaflow configuration.
+    node_selector: Union[Dict[str,str], str], optional, default None
+        Kubernetes node selector(s) to apply to the pod running the task.
+        Can be passed in as a comma separated string of values e.g.
+        'kubernetes.io/os=linux,kubernetes.io/arch=amd64' or as a dictionary
+        {'kubernetes.io/os': 'linux', 'kubernetes.io/arch': 'amd64'}
+    namespace : str, default METAFLOW_KUBERNETES_NAMESPACE
+        Kubernetes namespace to use when launching pod in Kubernetes.
+    gpu : int, optional, default None
+        Number of GPUs required for this step. A value of zero implies that
+        the scheduled node should not have GPUs.
+    gpu_vendor : str, default KUBERNETES_GPU_VENDOR
+        The vendor of the GPUs to be used for this step.
+    tolerations : List[Dict[str,str]], default []
+        The default is extracted from METAFLOW_KUBERNETES_TOLERATIONS.
+        Kubernetes tolerations to use when launching pod in Kubernetes.
+    labels: Dict[str, str], default: METAFLOW_KUBERNETES_LABELS
+        Kubernetes labels to use when launching pod in Kubernetes.
+    annotations: Dict[str, str], default: METAFLOW_KUBERNETES_ANNOTATIONS
+        Kubernetes annotations to use when launching pod in Kubernetes.
+    use_tmpfs : bool, default False
+        This enables an explicit tmpfs mount for this step.
+    tmpfs_tempdir : bool, default True
+        sets METAFLOW_TEMPDIR to tmpfs_path if set for this step.
+    tmpfs_size : int, optional, default: None
+        The value for the size (in MiB) of the tmpfs mount for this step.
+        This parameter maps to the `--tmpfs` option in Docker. Defaults to 50% of the
+        memory allocated for this step.
+    tmpfs_path : str, optional, default /metaflow_temp
+        Path to tmpfs mount for this step.
+    persistent_volume_claims : Dict[str, str], optional, default None
+        A map (dictionary) of persistent volumes to be mounted to the pod for this step. The map is from persistent
+        volumes to the path to which the volume is to be mounted, e.g., `{'pvc-name': '/path/to/mount/on'}`.
+    shared_memory: int, optional
+        Shared memory size (in MiB) required for this step
+    port: int, optional
+        Port number to specify in the Kubernetes job object
+    compute_pool : str, optional, default None
+        Compute pool to be used for for this step.
+        If not specified, any accessible compute pool within the perimeter is used.
+    hostname_resolution_timeout: int, default 10 * 60
+        Timeout in seconds for the workers tasks in the gang scheduled cluster to resolve the hostname of control task.
+        Only applicable when @parallel is used.
+    qos: str, default: Burstable
+        Quality of Service class to assign to the pod. Supported values are: Guaranteed, Burstable, BestEffort
+    
+    security_context: Dict[str, Any], optional, default None
+        Container security context. Applies to the task container. Allows the following keys:
+        - privileged: bool, optional, default None
+        - allow_privilege_escalation: bool, optional, default None
+        - run_as_user: int, optional, default None
+        - run_as_group: int, optional, default None
+        - run_as_non_root: bool, optional, default None
     """
     ...
 
-@typing.overload
-def timeout(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    ...
-
-@typing.overload
-def timeout(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def timeout(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, seconds: int = 0, minutes: int = 0, hours: int = 0):
+def ollama(*, models: list, backend: str, force_pull: bool, cache_update_policy: str, force_cache_update: bool, debug: bool, circuit_breaker_config: dict, timeout_config: dict) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies a timeout for your step.
+    This decorator is used to run Ollama APIs as Metaflow task sidecars.
     
-    This decorator is useful if this step may hang indefinitely.
+    User code call
+    --------------
+    @ollama(
+        models=[...],
+        ...
+    )
     
-    This can be used in conjunction with the `@retry` decorator as well as the `@catch` decorator.
-    A timeout is considered to be an exception thrown by the step. It will cause the step to be
-    retried if needed and the exception will be caught by the `@catch` decorator, if present.
+    Valid backend options
+    ---------------------
+    - 'local': Run as a separate process on the local task machine.
+    - (TODO) 'managed': Outerbounds hosts and selects compute provider.
+    - (TODO) 'remote': Spin up separate instance to serve Ollama models.
     
-    Note that all the values specified in parameters are added together so if you specify
-    60 seconds and 1 hour, the decorator will have an effective timeout of 1 hour and 1 minute.
+    Valid model options
+    -------------------
+    Any model here https://ollama.com/search, e.g. 'llama3.2', 'llama3.3'
     
     
     Parameters
     ----------
-    seconds : int, default 0
-        Number of seconds to wait prior to timing out.
-    minutes : int, default 0
-        Number of minutes to wait prior to timing out.
-    hours : int, default 0
-        Number of hours to wait prior to timing out.
+    models: list[str]
+        List of Ollama containers running models in sidecars.
+    backend: str
+        Determines where and how to run the Ollama process.
+    force_pull: bool
+        Whether to run `ollama pull` no matter what, or first check the remote cache in Metaflow datastore for this model key.
+    cache_update_policy: str
+        Cache update policy: "auto", "force", or "never".
+    force_cache_update: bool
+        Simple override for "force" cache update policy.
+    debug: bool
+        Whether to turn on verbose debugging logs.
+    circuit_breaker_config: dict
+        Configuration for circuit breaker protection. Keys: failure_threshold, recovery_timeout, reset_timeout.
+    timeout_config: dict
+        Configuration for various operation timeouts. Keys: pull, stop, health_check, install, server_startup.
     """
     ...
 
 @typing.overload
-def conda(*, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def parallel(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     """
-    Specifies the Conda environment for the step.
-    
-    Information in this decorator will augment any
-    attributes set in the `@conda_base` flow-level decorator. Hence,
-    you can use `@conda_base` to set packages required by all
-    steps and use `@conda` to specify step-specific overrides.
-    
-    
-    Parameters
-    ----------
-    packages : Dict[str, str], default {}
-        Packages to use for this step. The key is the name of the package
-        and the value is the version to use.
-    libraries : Dict[str, str], default {}
-        Supported for backward compatibility. When used with packages, packages will take precedence.
-    python : str, optional, default None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
-    disabled : bool, default False
-        If set to True, disables @conda.
+    Decorator prototype for all step decorators. This function gets specialized
+    and imported for all decorators types by _import_plugin_decorators().
     """
     ...
 
 @typing.overload
-def conda(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def parallel(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-@typing.overload
-def conda(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def parallel(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
+    """
+    Decorator prototype for all step decorators. This function gets specialized
+    and imported for all decorators types by _import_plugin_decorators().
+    """
     ...
 
-def conda(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False):
+def nvidia(*, gpu: int, gpu_type: str, queue_timeout: int) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies the Conda environment for the step.
-    
-    Information in this decorator will augment any
-    attributes set in the `@conda_base` flow-level decorator. Hence,
-    you can use `@conda_base` to set packages required by all
-    steps and use `@conda` to specify step-specific overrides.
+    Specifies that this step should execute on DGX cloud.
     
     
     Parameters
     ----------
-    packages : Dict[str, str], default {}
-        Packages to use for this step. The key is the name of the package
-        and the value is the version to use.
-    libraries : Dict[str, str], default {}
-        Supported for backward compatibility. When used with packages, packages will take precedence.
-    python : str, optional, default None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
-    disabled : bool, default False
-        If set to True, disables @conda.
+    gpu : int
+        Number of GPUs to use.
+    gpu_type : str
+        Type of Nvidia GPU to use.
+    queue_timeout : int
+        Time to keep the job in NVCF's queue.
     """
     ...
 
-@typing.overload
-def resources(*, cpu: int = 1, gpu: typing.Optional[int] = None, disk: typing.Optional[int] = None, memory: int = 4096, shared_memory: typing.Optional[int] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def nvct(*, gpu: int, gpu_type: str) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies the resources needed when executing this step.
-    
-    Use `@resources` to specify the resource requirements
-    independently of the specific compute layer (`@batch`, `@kubernetes`).
-    
-    You can choose the compute layer on the command line by executing e.g.
-    ```
-    python myflow.py run --with batch
-    ```
-    or
-    ```
-    python myflow.py run --with kubernetes
-    ```
-    which executes the flow on the desired system using the
-    requirements specified in `@resources`.
+    Specifies that this step should execute on DGX cloud.
     
     
     Parameters
     ----------
-    cpu : int, default 1
-        Number of CPUs required for this step.
-    gpu : int, optional, default None
-        Number of GPUs required for this step.
-    disk : int, optional, default None
-        Disk size (in MB) required for this step. Only applies on Kubernetes.
-    memory : int, default 4096
-        Memory size (in MB) required for this step.
-    shared_memory : int, optional, default None
-        The value for the size (in MiB) of the /dev/shm volume for this step.
-        This parameter maps to the `--shm-size` option in Docker.
+    gpu : int
+        Number of GPUs to use.
+    gpu_type : str
+        Type of Nvidia GPU to use.
     """
     ...
 
 @typing.overload
-def resources(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    ...
-
-@typing.overload
-def resources(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def resources(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, cpu: int = 1, gpu: typing.Optional[int] = None, disk: typing.Optional[int] = None, memory: int = 4096, shared_memory: typing.Optional[int] = None):
+def environment(*, vars: typing.Dict[str, str] = {}) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies the resources needed when executing this step.
-    
-    Use `@resources` to specify the resource requirements
-    independently of the specific compute layer (`@batch`, `@kubernetes`).
-    
-    You can choose the compute layer on the command line by executing e.g.
-    ```
-    python myflow.py run --with batch
-    ```
-    or
-    ```
-    python myflow.py run --with kubernetes
-    ```
-    which executes the flow on the desired system using the
-    requirements specified in `@resources`.
+    Specifies environment variables to be set prior to the execution of a step.
     
     
     Parameters
     ----------
-    cpu : int, default 1
-        Number of CPUs required for this step.
-    gpu : int, optional, default None
-        Number of GPUs required for this step.
-    disk : int, optional, default None
-        Disk size (in MB) required for this step. Only applies on Kubernetes.
-    memory : int, default 4096
-        Memory size (in MB) required for this step.
-    shared_memory : int, optional, default None
-        The value for the size (in MiB) of the /dev/shm volume for this step.
-        This parameter maps to the `--shm-size` option in Docker.
+    vars : Dict[str, str], default {}
+        Dictionary of environment variables to set.
     """
     ...
 
 @typing.overload
-def test_append_card(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    """
-    A simple decorator that demonstrates using CardDecoratorInjector
-    to inject a card and render simple markdown content.
-    """
+def environment(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def test_append_card(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
-    ...
-
-def test_append_card(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None):
-    """
-    A simple decorator that demonstrates using CardDecoratorInjector
-    to inject a card and render simple markdown content.
-    """
+def environment(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def nvidia(*, gpu: int, gpu_type: str, queue_timeout: int) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def environment(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, vars: typing.Dict[str, str] = {}):
     """
-    Specifies that this step should execute on DGX cloud.
+    Specifies environment variables to be set prior to the execution of a step.
     
     
     Parameters
     ----------
-    gpu : int
-        Number of GPUs to use.
-    gpu_type : str
-        Type of Nvidia GPU to use.
-    queue_timeout : int
-        Time to keep the job in NVCF's queue.
+    vars : Dict[str, str], default {}
+        Dictionary of environment variables to set.
     """
     ...
 
 @typing.overload
-def secrets(*, sources: typing.List[typing.Union[str, typing.Dict[str, typing.Any]]] = [], role: typing.Optional[str] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def timeout(*, seconds: int = 0, minutes: int = 0, hours: int = 0) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies secrets to be retrieved and injected as environment variables prior to
-    the execution of a step.
+    Specifies a timeout for your step.
+    
+    This decorator is useful if this step may hang indefinitely.
+    
+    This can be used in conjunction with the `@retry` decorator as well as the `@catch` decorator.
+    A timeout is considered to be an exception thrown by the step. It will cause the step to be
+    retried if needed and the exception will be caught by the `@catch` decorator, if present.
+    
+    Note that all the values specified in parameters are added together so if you specify
+    60 seconds and 1 hour, the decorator will have an effective timeout of 1 hour and 1 minute.
     
     
     Parameters
     ----------
-    sources : List[Union[str, Dict[str, Any]]], default: []
-        List of secret specs, defining how the secrets are to be retrieved
-    role : str, optional, default: None
-        Role to use for fetching secrets
+    seconds : int, default 0
+        Number of seconds to wait prior to timing out.
+    minutes : int, default 0
+        Number of minutes to wait prior to timing out.
+    hours : int, default 0
+        Number of hours to wait prior to timing out.
     """
     ...
 
 @typing.overload
-def secrets(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def timeout(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def secrets(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def timeout(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def secrets(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, sources: typing.List[typing.Union[str, typing.Dict[str, typing.Any]]] = [], role: typing.Optional[str] = None):
+def timeout(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, seconds: int = 0, minutes: int = 0, hours: int = 0):
     """
-    Specifies secrets to be retrieved and injected as environment variables prior to
-    the execution of a step.
+    Specifies a timeout for your step.
+    
+    This decorator is useful if this step may hang indefinitely.
+    
+    This can be used in conjunction with the `@retry` decorator as well as the `@catch` decorator.
+    A timeout is considered to be an exception thrown by the step. It will cause the step to be
+    retried if needed and the exception will be caught by the `@catch` decorator, if present.
+    
+    Note that all the values specified in parameters are added together so if you specify
+    60 seconds and 1 hour, the decorator will have an effective timeout of 1 hour and 1 minute.
     
     
     Parameters
     ----------
-    sources : List[Union[str, Dict[str, Any]]], default: []
-        List of secret specs, defining how the secrets are to be retrieved
-    role : str, optional, default: None
-        Role to use for fetching secrets
+    seconds : int, default 0
+        Number of seconds to wait prior to timing out.
+    minutes : int, default 0
+        Number of minutes to wait prior to timing out.
+    hours : int, default 0
+        Number of hours to wait prior to timing out.
     """
     ...
 
-def coreweave_s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def s3_proxy(*, integration_name: typing.Optional[str] = None, write_mode: typing.Optional[str] = None, debug: typing.Optional[bool] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    `@coreweave_s3_proxy` is a CoreWeave-specific S3 Proxy decorator for routing S3 requests through a local proxy service.
-    It exists to make it easier for users to know that this decorator should only be used with
-    a Neo Cloud like CoreWeave. The underlying mechanics of the decorator is the same as the `@s3_proxy`:
-    
-    
     Set up an S3 proxy that caches objects in an external, S3‑compatible bucket
     for S3 read and write requests.
     
@@ -1376,251 +1226,338 @@ def coreweave_s3_proxy(*, integration_name: typing.Optional[str] = None, write_m
     ...
 
 @typing.overload
-def environment(*, vars: typing.Dict[str, str] = {}) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def secrets(*, sources: typing.List[typing.Union[str, typing.Dict[str, typing.Any]]] = [], role: typing.Optional[str] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies environment variables to be set prior to the execution of a step.
+    Specifies secrets to be retrieved and injected as environment variables prior to
+    the execution of a step.
     
     
     Parameters
     ----------
-    vars : Dict[str, str], default {}
-        Dictionary of environment variables to set.
+    sources : List[Union[str, Dict[str, Any]]], default: []
+        List of secret specs, defining how the secrets are to be retrieved
+    role : str, optional, default: None
+        Role to use for fetching secrets
     """
     ...
 
 @typing.overload
-def environment(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def secrets(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def environment(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def secrets(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def environment(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, vars: typing.Dict[str, str] = {}):
+def secrets(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, sources: typing.List[typing.Union[str, typing.Dict[str, typing.Any]]] = [], role: typing.Optional[str] = None):
     """
-    Specifies environment variables to be set prior to the execution of a step.
+    Specifies secrets to be retrieved and injected as environment variables prior to
+    the execution of a step.
     
     
     Parameters
     ----------
-    vars : Dict[str, str], default {}
-        Dictionary of environment variables to set.
+    sources : List[Union[str, Dict[str, Any]]], default: []
+        List of secret specs, defining how the secrets are to be retrieved
+    role : str, optional, default: None
+        Role to use for fetching secrets
     """
     ...
 
 @typing.overload
-def pypi(*, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def resources(*, cpu: int = 1, gpu: typing.Optional[int] = None, disk: typing.Optional[int] = None, memory: int = 4096, shared_memory: typing.Optional[int] = None) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
     """
-    Specifies the PyPI packages for the step.
+    Specifies the resources needed when executing this step.
     
-    Information in this decorator will augment any
-    attributes set in the `@pyi_base` flow-level decorator. Hence,
-    you can use `@pypi_base` to set packages required by all
-    steps and use `@pypi` to specify step-specific overrides.
+    Use `@resources` to specify the resource requirements
+    independently of the specific compute layer (`@batch`, `@kubernetes`).
+    
+    You can choose the compute layer on the command line by executing e.g.
+    ```
+    python myflow.py run --with batch
+    ```
+    or
+    ```
+    python myflow.py run --with kubernetes
+    ```
+    which executes the flow on the desired system using the
+    requirements specified in `@resources`.
     
     
     Parameters
     ----------
-    packages : Dict[str, str], default: {}
-        Packages to use for this step. The key is the name of the package
-        and the value is the version to use.
-    python : str, optional, default: None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
+    cpu : int, default 1
+        Number of CPUs required for this step.
+    gpu : int, optional, default None
+        Number of GPUs required for this step.
+    disk : int, optional, default None
+        Disk size (in MB) required for this step. Only applies on Kubernetes.
+    memory : int, default 4096
+        Memory size (in MB) required for this step.
+    shared_memory : int, optional, default None
+        The value for the size (in MiB) of the /dev/shm volume for this step.
+        This parameter maps to the `--shm-size` option in Docker.
     """
     ...
 
 @typing.overload
-def pypi(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
+def resources(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
     ...
 
 @typing.overload
-def pypi(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def resources(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
     ...
 
-def pypi(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None):
+def resources(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, cpu: int = 1, gpu: typing.Optional[int] = None, disk: typing.Optional[int] = None, memory: int = 4096, shared_memory: typing.Optional[int] = None):
     """
-    Specifies the PyPI packages for the step.
+    Specifies the resources needed when executing this step.
+    
+    Use `@resources` to specify the resource requirements
+    independently of the specific compute layer (`@batch`, `@kubernetes`).
+    
+    You can choose the compute layer on the command line by executing e.g.
+    ```
+    python myflow.py run --with batch
+    ```
+    or
+    ```
+    python myflow.py run --with kubernetes
+    ```
+    which executes the flow on the desired system using the
+    requirements specified in `@resources`.
+    
+    
+    Parameters
+    ----------
+    cpu : int, default 1
+        Number of CPUs required for this step.
+    gpu : int, optional, default None
+        Number of GPUs required for this step.
+    disk : int, optional, default None
+        Disk size (in MB) required for this step. Only applies on Kubernetes.
+    memory : int, default 4096
+        Memory size (in MB) required for this step.
+    shared_memory : int, optional, default None
+        The value for the size (in MiB) of the /dev/shm volume for this step.
+        This parameter maps to the `--shm-size` option in Docker.
+    """
+    ...
+
+def vllm(*, model: str, backend: str, openai_api_server: bool, debug: bool, card_refresh_interval: int, max_retries: int, retry_alert_frequency: int, engine_args: dict) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+    """
+    This decorator is used to run vllm APIs as Metaflow task sidecars.
+    
+    User code call
+    --------------
+    @vllm(
+        model="...",
+        ...
+    )
+    
+    Valid backend options
+    ---------------------
+    - 'local': Run as a separate process on the local task machine.
+    
+    Valid model options
+    -------------------
+    Any HuggingFace model identifier, e.g. 'meta-llama/Llama-3.2-1B'
+    
+    NOTE: vLLM's OpenAI-compatible server serves ONE model per server instance.
+    If you need multiple models, you must create multiple @vllm decorators.
+    
+    
+    Parameters
+    ----------
+    model: str
+        HuggingFace model identifier to be served by vLLM.
+    backend: str
+        Determines where and how to run the vLLM process.
+    openai_api_server: bool
+        Whether to use OpenAI-compatible API server mode (subprocess) instead of native engine.
+        Default is False (uses native engine).
+        Set to True for backward compatibility with existing code.
+    debug: bool
+        Whether to turn on verbose debugging logs.
+    card_refresh_interval: int
+        Interval in seconds for refreshing the vLLM status card.
+        Only used when openai_api_server=True.
+    max_retries: int
+        Maximum number of retries checking for vLLM server startup.
+        Only used when openai_api_server=True.
+    retry_alert_frequency: int
+        Frequency of alert logs for vLLM server startup retries.
+        Only used when openai_api_server=True.
+    engine_args : dict
+        Additional keyword arguments to pass to the vLLM engine.
+        For example, `tensor_parallel_size=2`.
+    """
+    ...
+
+def huggingface_hub(*, temp_dir_root: typing.Optional[str] = None, cache_scope: typing.Optional[str] = None, load: typing.Union[typing.List[str], typing.List[typing.Tuple[typing.Dict, str]], typing.List[typing.Tuple[str, str]], typing.List[typing.Dict], None]) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+    """
+    Decorator that helps cache, version, and store models/datasets from the Hugging Face Hub.
+    
+    Examples
+    --------
+    
+    ```python
+    # **Usage: creating references to models from the Hugging Face Hub that may be loaded in downstream steps**
+    @huggingface_hub
+    @step
+    def pull_model_from_huggingface(self):
+        # `current.huggingface_hub.snapshot_download` downloads the model from the Hugging Face Hub
+        # and saves it in the backend storage based on the model's `repo_id`. If there exists a model
+        # with the same `repo_id` in the backend storage, it will not download the model again. The return
+        # value of the function is a reference to the model in the backend storage.
+        # This reference can be used to load the model in the subsequent steps via `@model(load=["llama_model"])`
+    
+        self.model_id = "mistralai/Mistral-7B-Instruct-v0.1"
+        self.llama_model = current.huggingface_hub.snapshot_download(
+            repo_id=self.model_id,
+            allow_patterns=["*.safetensors", "*.json", "tokenizer.*"],
+        )
+        self.next(self.train)
+    
+    # **Usage: explicitly loading models at runtime from the Hugging Face Hub or from cache (from Metaflow's datastore)**
+    @huggingface_hub
+    @step
+    def run_training(self):
+        # Temporary directory (auto-cleaned on exit)
+        with current.huggingface_hub.load(
+            repo_id="google-bert/bert-base-uncased",
+            allow_patterns=["*.bin"],
+        ) as local_path:
+            # Use files under local_path
+            train_model(local_path)
+            ...
+    
+    # **Usage: loading models directly from the Hugging Face Hub or from cache (from Metaflow's datastore)**
+    
+    @huggingface_hub(load=["mistralai/Mistral-7B-Instruct-v0.1"])
+    @step
+    def pull_model_from_huggingface(self):
+        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
+    
+    @huggingface_hub(load=[("mistralai/Mistral-7B-Instruct-v0.1", "/my-directory"), ("myorg/mistral-lora", "/my-lora-directory")])
+    @step
+    def finetune_model(self):
+        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
+        # path_to_model will be /my-directory
+    
+    
+    # Takes all the arguments passed to `snapshot_download`
+    # except for `local_dir`
+    @huggingface_hub(load=[
+        {
+            "repo_id": "mistralai/Mistral-7B-Instruct-v0.1",
+        },
+        {
+            "repo_id": "myorg/mistral-lora",
+            "repo_type": "model",
+        },
+    ])
+    @step
+    def finetune_model(self):
+        path_to_model = current.huggingface_hub.loaded["mistralai/Mistral-7B-Instruct-v0.1"]
+        # path_to_model will be /my-directory
+    ```
+    
+    
+    Parameters
+    ----------
+    temp_dir_root : str, optional
+        The root directory that will hold the temporary directory where objects will be downloaded.
+    
+    cache_scope : str, optional
+        The scope of the cache. Can be `checkpoint` / `flow` / `global`.
+            - `checkpoint` (default): All repos are stored like objects saved by `@checkpoint`.
+                i.e., the cached path is derived from the namespace, flow, step, and Metaflow foreach iteration.
+                Any repo downloaded under this scope will only be retrieved from the cache when the step runs under the same namespace in the same flow (at the same foreach index).
+    
+            - `flow`: All repos are cached under the flow, regardless of namespace.
+                i.e., the cached path is derived solely from the flow name.
+                When to use this mode: (1) Multiple users are executing the same flow and want shared access to the repos cached by the decorator. (2) Multiple versions of a flow are deployed, all needing access to the same repos cached by the decorator.
+    
+            - `global`: All repos are cached under a globally static path.
+                i.e., the base path of the cache is static and all repos are stored under it.
+                When to use this mode:
+                    - All repos from the Hugging Face Hub need to be shared by users across all flow executions.
+            - Each caching scope comes with its own trade-offs:
+                - `checkpoint`:
+                    - Has explicit control over when caches are populated (controlled by the same flow that has the `@huggingface_hub` decorator) but ends up hitting the Hugging Face Hub more often if there are many users/namespaces/steps.
+                    - Since objects are written on a `namespace/flow/step` basis, the blast radius of a bad checkpoint is limited to a particular flow in a namespace.
+                - `flow`:
+                    - Has less control over when caches are populated (can be written by any execution instance of a flow from any namespace) but results in more cache hits.
+                    - The blast radius of a bad checkpoint is limited to all runs of a particular flow.
+                    - It doesn't promote cache reuse across flows.
+                - `global`:
+                    - Has no control over when caches are populated (can be written by any flow execution) but has the highest cache hit rate.
+                    - It promotes cache reuse across flows.
+                    - The blast radius of a bad checkpoint spans every flow that could be using a particular repo.
     
-    Information in this decorator will augment any
-    attributes set in the `@pyi_base` flow-level decorator. Hence,
-    you can use `@pypi_base` to set packages required by all
-    steps and use `@pypi` to specify step-specific overrides.
+    load: Union[List[str], List[Tuple[Dict, str]], List[Tuple[str, str]], List[Dict], None]
+        The list of repos (models/datasets) to load.
     
+        Loaded repos can be accessed via `current.huggingface_hub.loaded`. If load is set, then the following happens:
     
-    Parameters
-    ----------
-    packages : Dict[str, str], default: {}
-        Packages to use for this step. The key is the name of the package
-        and the value is the version to use.
-    python : str, optional, default: None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
+        - If repo (model/dataset) is not found in the datastore:
+            - Downloads the repo from Hugging Face Hub to a temporary directory (or uses specified path) for local access
+            - Stores it in Metaflow's datastore (s3/gcs/azure etc.) with a unique name based on repo_type/repo_id
+                - All HF models loaded for a `@step` will be cached separately under flow/step/namespace.
+    
+        - If repo is found in the datastore:
+            - Loads it directly from datastore to local path (can be temporary directory or specified path)
     """
     ...
 
 @typing.overload
-def catch(*, var: typing.Optional[str] = None, print_exception: bool = True) -> typing.Callable[[typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]], typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]]]:
+def schedule(*, hourly: bool = False, daily: bool = True, weekly: bool = False, cron: typing.Optional[str] = None, timezone: typing.Optional[str] = None) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
     """
-    Specifies that the step will success under all circumstances.
-    
-    The decorator will create an optional artifact, specified by `var`, which
-    contains the exception raised. You can use it to detect the presence
-    of errors, indicating that all happy-path artifacts produced by the step
-    are missing.
+    Specifies the times when the flow should be run when running on a
+    production scheduler.
     
     
     Parameters
     ----------
-    var : str, optional, default None
-        Name of the artifact in which to store the caught exception.
-        If not specified, the exception is not stored.
-    print_exception : bool, default True
-        Determines whether or not the exception is printed to
-        stdout when caught.
+    hourly : bool, default False
+        Run the workflow hourly.
+    daily : bool, default True
+        Run the workflow daily.
+    weekly : bool, default False
+        Run the workflow weekly.
+    cron : str, optional, default None
+        Run the workflow at [a custom Cron schedule](https://docs.aws.amazon.com/eventbridge/latest/userguide/scheduled-events.html#cron-expressions)
+        specified by this expression.
+    timezone : str, optional, default None
+        Timezone on which the schedule runs (default: None). Currently supported only for Argo workflows,
+        which accepts timezones in [IANA format](https://nodatime.org/TimeZones).
     """
     ...
 
 @typing.overload
-def catch(f: typing.Callable[[FlowSpecDerived, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, StepFlag], None]:
-    ...
-
-@typing.overload
-def catch(f: typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]) -> typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None]:
+def schedule(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
     ...
 
-def catch(f: typing.Union[typing.Callable[[FlowSpecDerived, StepFlag], None], typing.Callable[[FlowSpecDerived, typing.Any, StepFlag], None], None] = None, *, var: typing.Optional[str] = None, print_exception: bool = True):
+def schedule(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, hourly: bool = False, daily: bool = True, weekly: bool = False, cron: typing.Optional[str] = None, timezone: typing.Optional[str] = None):
     """
-    Specifies that the step will success under all circumstances.
-    
-    The decorator will create an optional artifact, specified by `var`, which
-    contains the exception raised. You can use it to detect the presence
-    of errors, indicating that all happy-path artifacts produced by the step
-    are missing.
+    Specifies the times when the flow should be run when running on a
+    production scheduler.
     
     
     Parameters
     ----------
-    var : str, optional, default None
-        Name of the artifact in which to store the caught exception.
-        If not specified, the exception is not stored.
-    print_exception : bool, default True
-        Determines whether or not the exception is printed to
-        stdout when caught.
-    """
-    ...
-
-def with_artifact_store(f: typing.Optional[typing.Type[FlowSpecDerived]] = None):
-    """
-    Allows setting external datastores to save data for the
-    `@checkpoint`/`@model`/`@huggingface_hub` decorators.
-    
-    This decorator is useful when users wish to save data to a different datastore
-    than what is configured in Metaflow. This can be for variety of reasons:
-    
-    1. Data security: The objects needs to be stored in a bucket (object storage) that is not accessible by other flows.
-    2. Data Locality: The location where the task is executing is not located in the same region as the datastore.
-        - Example: Metaflow datastore lives in US East, but the task is executing in Finland datacenters.
-    3. Data Lifecycle Policies: The objects need to be archived / managed separately from the Metaflow managed objects.
-        - Example: Flow is training very large models that need to be stored separately and will be deleted more aggressively than the Metaflow managed objects.
-    
-    Usage:
-    ----------
-    
-    - Using a custom IAM role to access the datastore.
-    
-        ```python
-        @with_artifact_store(
-            type="s3",
-            config=lambda: {
-                "root": "s3://my-bucket-foo/path/to/root",
-                "role_arn": ROLE,
-            },
-        )
-        class MyFlow(FlowSpec):
-    
-            @checkpoint
-            @step
-            def start(self):
-                with open("my_file.txt", "w") as f:
-                    f.write("Hello, World!")
-                self.external_bucket_checkpoint = current.checkpoint.save("my_file.txt")
-                self.next(self.end)
-    
-        ```
-    
-    - Using credentials to access the s3-compatible datastore.
-    
-        ```python
-        @with_artifact_store(
-            type="s3",
-            config=lambda: {
-                "root": "s3://my-bucket-foo/path/to/root",
-                "client_params": {
-                    "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
-                    "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
-                },
-            },
-        )
-        class MyFlow(FlowSpec):
-    
-            @checkpoint
-            @step
-            def start(self):
-                with open("my_file.txt", "w") as f:
-                    f.write("Hello, World!")
-                self.external_bucket_checkpoint = current.checkpoint.save("my_file.txt")
-                self.next(self.end)
-    
-        ```
-    
-    - Accessing objects stored in external datastores after task execution.
-    
-        ```python
-        run = Run("CheckpointsTestsFlow/8992")
-        with artifact_store_from(run=run, config={
-            "client_params": {
-                "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
-                "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
-            },
-        }):
-            with Checkpoint() as cp:
-                latest = cp.list(
-                    task=run["start"].task
-                )[0]
-                print(latest)
-                cp.load(
-                    latest,
-                    "test-checkpoints"
-                )
-    
-        task = Task("TorchTuneFlow/8484/train/53673")
-        with artifact_store_from(run=run, config={
-            "client_params": {
-                "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
-                "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
-            },
-        }):
-            load_model(
-                task.data.model_ref,
-                "test-models"
-            )
-        ```
-    Parameters:
-    ----------
-    
-    type: str
-        The type of the datastore. Can be one of 's3', 'gcs', 'azure' or any other supported metaflow Datastore.
-    
-    config: dict or Callable
-        Dictionary of configuration options for the datastore. The following keys are required:
-        - root: The root path in the datastore where the data will be saved. (needs to be in the format expected by the datastore)
-            - example: 's3://bucket-name/path/to/root'
-            - example: 'gs://bucket-name/path/to/root'
-            - example: 'https://myblockacc.blob.core.windows.net/metaflow/'
-        - role_arn (optional): AWS IAM role to access s3 bucket (only when `type` is 's3')
-        - session_vars (optional): AWS session variables to access s3 bucket (only when `type` is 's3')
-        - client_params (optional): AWS client parameters to access s3 bucket (only when `type` is 's3')
+    hourly : bool, default False
+        Run the workflow hourly.
+    daily : bool, default True
+        Run the workflow daily.
+    weekly : bool, default False
+        Run the workflow weekly.
+    cron : str, optional, default None
+        Run the workflow at [a custom Cron schedule](https://docs.aws.amazon.com/eventbridge/latest/userguide/scheduled-events.html#cron-expressions)
+        specified by this expression.
+    timezone : str, optional, default None
+        Timezone on which the schedule runs (default: None). Currently supported only for Argo workflows,
+        which accepts timezones in [IANA format](https://nodatime.org/TimeZones).
     """
     ...
 
@@ -1652,59 +1589,18 @@ def airflow_external_task_sensor(*, timeout: int, poke_interval: int, mode: str,
     external_dag_id : str
         The dag_id that contains the task you want to wait for.
     external_task_ids : List[str]
-        The list of task_ids that you want to wait for.
-        If None (default value) the sensor waits for the DAG. (Default: None)
-    allowed_states : List[str]
-        Iterable of allowed states, (Default: ['success'])
-    failed_states : List[str]
-        Iterable of failed or dis-allowed states. (Default: None)
-    execution_delta : datetime.timedelta
-        time difference with the previous execution to look at,
-        the default is the same logical date as the current task or DAG. (Default: None)
-    check_existence: bool
-        Set to True to check if the external task exists or check if
-        the DAG to wait for exists. (Default: True)
-    """
-    ...
-
-@typing.overload
-def pypi_base(*, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
-    """
-    Specifies the PyPI packages for all steps of the flow.
-    
-    Use `@pypi_base` to set common packages required by all
-    steps and use `@pypi` to specify step-specific overrides.
-    
-    Parameters
-    ----------
-    packages : Dict[str, str], default: {}
-        Packages to use for this flow. The key is the name of the package
-        and the value is the version to use.
-    python : str, optional, default: None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
-    """
-    ...
-
-@typing.overload
-def pypi_base(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
-    ...
-
-def pypi_base(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None):
-    """
-    Specifies the PyPI packages for all steps of the flow.
-    
-    Use `@pypi_base` to set common packages required by all
-    steps and use `@pypi` to specify step-specific overrides.
-    
-    Parameters
-    ----------
-    packages : Dict[str, str], default: {}
-        Packages to use for this flow. The key is the name of the package
-        and the value is the version to use.
-    python : str, optional, default: None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
+        The list of task_ids that you want to wait for.
+        If None (default value) the sensor waits for the DAG. (Default: None)
+    allowed_states : List[str]
+        Iterable of allowed states, (Default: ['success'])
+    failed_states : List[str]
+        Iterable of failed or dis-allowed states. (Default: None)
+    execution_delta : datetime.timedelta
+        time difference with the previous execution to look at,
+        the default is the same logical date as the current task or DAG. (Default: None)
+    check_existence: bool
+        Set to True to check if the external task exists or check if
+        the DAG to wait for exists. (Default: True)
     """
     ...
 
@@ -1751,6 +1647,47 @@ def airflow_s3_key_sensor(*, timeout: int, poke_interval: int, mode: str, expone
     """
     ...
 
+@typing.overload
+def pypi_base(*, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
+    """
+    Specifies the PyPI packages for all steps of the flow.
+    
+    Use `@pypi_base` to set common packages required by all
+    steps and use `@pypi` to specify step-specific overrides.
+    
+    Parameters
+    ----------
+    packages : Dict[str, str], default: {}
+        Packages to use for this flow. The key is the name of the package
+        and the value is the version to use.
+    python : str, optional, default: None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    """
+    ...
+
+@typing.overload
+def pypi_base(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
+    ...
+
+def pypi_base(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, packages: typing.Dict[str, str] = {}, python: typing.Optional[str] = None):
+    """
+    Specifies the PyPI packages for all steps of the flow.
+    
+    Use `@pypi_base` to set common packages required by all
+    steps and use `@pypi` to specify step-specific overrides.
+    
+    Parameters
+    ----------
+    packages : Dict[str, str], default: {}
+        Packages to use for this flow. The key is the name of the package
+        and the value is the version to use.
+    python : str, optional, default: None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    """
+    ...
+
 def project(*, name: str, branch: typing.Optional[str] = None, production: bool = False) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
     """
     Specifies what flows belong to the same project.
@@ -1787,53 +1724,53 @@ def project(*, name: str, branch: typing.Optional[str] = None, production: bool
     ...
 
 @typing.overload
-def schedule(*, hourly: bool = False, daily: bool = True, weekly: bool = False, cron: typing.Optional[str] = None, timezone: typing.Optional[str] = None) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
+def conda_base(*, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
     """
-    Specifies the times when the flow should be run when running on a
-    production scheduler.
+    Specifies the Conda environment for all steps of the flow.
+    
+    Use `@conda_base` to set common libraries required by all
+    steps and use `@conda` to specify step-specific additions.
     
     
     Parameters
     ----------
-    hourly : bool, default False
-        Run the workflow hourly.
-    daily : bool, default True
-        Run the workflow daily.
-    weekly : bool, default False
-        Run the workflow weekly.
-    cron : str, optional, default None
-        Run the workflow at [a custom Cron schedule](https://docs.aws.amazon.com/eventbridge/latest/userguide/scheduled-events.html#cron-expressions)
-        specified by this expression.
-    timezone : str, optional, default None
-        Timezone on which the schedule runs (default: None). Currently supported only for Argo workflows,
-        which accepts timezones in [IANA format](https://nodatime.org/TimeZones).
+    packages : Dict[str, str], default {}
+        Packages to use for this flow. The key is the name of the package
+        and the value is the version to use.
+    libraries : Dict[str, str], default {}
+        Supported for backward compatibility. When used with packages, packages will take precedence.
+    python : str, optional, default None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    disabled : bool, default False
+        If set to True, disables Conda.
     """
     ...
 
 @typing.overload
-def schedule(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
+def conda_base(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
     ...
 
-def schedule(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, hourly: bool = False, daily: bool = True, weekly: bool = False, cron: typing.Optional[str] = None, timezone: typing.Optional[str] = None):
+def conda_base(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False):
     """
-    Specifies the times when the flow should be run when running on a
-    production scheduler.
+    Specifies the Conda environment for all steps of the flow.
+    
+    Use `@conda_base` to set common libraries required by all
+    steps and use `@conda` to specify step-specific additions.
     
     
     Parameters
     ----------
-    hourly : bool, default False
-        Run the workflow hourly.
-    daily : bool, default True
-        Run the workflow daily.
-    weekly : bool, default False
-        Run the workflow weekly.
-    cron : str, optional, default None
-        Run the workflow at [a custom Cron schedule](https://docs.aws.amazon.com/eventbridge/latest/userguide/scheduled-events.html#cron-expressions)
-        specified by this expression.
-    timezone : str, optional, default None
-        Timezone on which the schedule runs (default: None). Currently supported only for Argo workflows,
-        which accepts timezones in [IANA format](https://nodatime.org/TimeZones).
+    packages : Dict[str, str], default {}
+        Packages to use for this flow. The key is the name of the package
+        and the value is the version to use.
+    libraries : Dict[str, str], default {}
+        Supported for backward compatibility. When used with packages, packages will take precedence.
+    python : str, optional, default None
+        Version of Python to use, e.g. '3.7.4'. A default value of None implies
+        that the version used will correspond to the version of the Python interpreter used to start the run.
+    disabled : bool, default False
+        If set to True, disables Conda.
     """
     ...
 
@@ -1930,54 +1867,117 @@ def trigger(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, event: t
     """
     ...
 
-@typing.overload
-def conda_base(*, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False) -> typing.Callable[[typing.Type[FlowSpecDerived]], typing.Type[FlowSpecDerived]]:
+def with_artifact_store(f: typing.Optional[typing.Type[FlowSpecDerived]] = None):
     """
-    Specifies the Conda environment for all steps of the flow.
+    Allows setting external datastores to save data for the
+    `@checkpoint`/`@model`/`@huggingface_hub` decorators.
     
-    Use `@conda_base` to set common libraries required by all
-    steps and use `@conda` to specify step-specific additions.
+    This decorator is useful when users wish to save data to a different datastore
+    than what is configured in Metaflow. This can be for variety of reasons:
     
+    1. Data security: The objects needs to be stored in a bucket (object storage) that is not accessible by other flows.
+    2. Data Locality: The location where the task is executing is not located in the same region as the datastore.
+        - Example: Metaflow datastore lives in US East, but the task is executing in Finland datacenters.
+    3. Data Lifecycle Policies: The objects need to be archived / managed separately from the Metaflow managed objects.
+        - Example: Flow is training very large models that need to be stored separately and will be deleted more aggressively than the Metaflow managed objects.
     
-    Parameters
+    Usage:
     ----------
-    packages : Dict[str, str], default {}
-        Packages to use for this flow. The key is the name of the package
-        and the value is the version to use.
-    libraries : Dict[str, str], default {}
-        Supported for backward compatibility. When used with packages, packages will take precedence.
-    python : str, optional, default None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
-    disabled : bool, default False
-        If set to True, disables Conda.
-    """
-    ...
-
-@typing.overload
-def conda_base(f: typing.Type[FlowSpecDerived]) -> typing.Type[FlowSpecDerived]:
-    ...
-
-def conda_base(f: typing.Optional[typing.Type[FlowSpecDerived]] = None, *, packages: typing.Dict[str, str] = {}, libraries: typing.Dict[str, str] = {}, python: typing.Optional[str] = None, disabled: bool = False):
-    """
-    Specifies the Conda environment for all steps of the flow.
     
-    Use `@conda_base` to set common libraries required by all
-    steps and use `@conda` to specify step-specific additions.
+    - Using a custom IAM role to access the datastore.
+    
+        ```python
+        @with_artifact_store(
+            type="s3",
+            config=lambda: {
+                "root": "s3://my-bucket-foo/path/to/root",
+                "role_arn": ROLE,
+            },
+        )
+        class MyFlow(FlowSpec):
+    
+            @checkpoint
+            @step
+            def start(self):
+                with open("my_file.txt", "w") as f:
+                    f.write("Hello, World!")
+                self.external_bucket_checkpoint = current.checkpoint.save("my_file.txt")
+                self.next(self.end)
     
+        ```
     
-    Parameters
+    - Using credentials to access the s3-compatible datastore.
+    
+        ```python
+        @with_artifact_store(
+            type="s3",
+            config=lambda: {
+                "root": "s3://my-bucket-foo/path/to/root",
+                "client_params": {
+                    "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
+                    "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
+                },
+            },
+        )
+        class MyFlow(FlowSpec):
+    
+            @checkpoint
+            @step
+            def start(self):
+                with open("my_file.txt", "w") as f:
+                    f.write("Hello, World!")
+                self.external_bucket_checkpoint = current.checkpoint.save("my_file.txt")
+                self.next(self.end)
+    
+        ```
+    
+    - Accessing objects stored in external datastores after task execution.
+    
+        ```python
+        run = Run("CheckpointsTestsFlow/8992")
+        with artifact_store_from(run=run, config={
+            "client_params": {
+                "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
+                "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
+            },
+        }):
+            with Checkpoint() as cp:
+                latest = cp.list(
+                    task=run["start"].task
+                )[0]
+                print(latest)
+                cp.load(
+                    latest,
+                    "test-checkpoints"
+                )
+    
+        task = Task("TorchTuneFlow/8484/train/53673")
+        with artifact_store_from(run=run, config={
+            "client_params": {
+                "aws_access_key_id": os.environ.get("MY_CUSTOM_ACCESS_KEY"),
+                "aws_secret_access_key": os.environ.get("MY_CUSTOM_SECRET_KEY"),
+            },
+        }):
+            load_model(
+                task.data.model_ref,
+                "test-models"
+            )
+        ```
+    Parameters:
     ----------
-    packages : Dict[str, str], default {}
-        Packages to use for this flow. The key is the name of the package
-        and the value is the version to use.
-    libraries : Dict[str, str], default {}
-        Supported for backward compatibility. When used with packages, packages will take precedence.
-    python : str, optional, default None
-        Version of Python to use, e.g. '3.7.4'. A default value of None implies
-        that the version used will correspond to the version of the Python interpreter used to start the run.
-    disabled : bool, default False
-        If set to True, disables Conda.
+    
+    type: str
+        The type of the datastore. Can be one of 's3', 'gcs', 'azure' or any other supported metaflow Datastore.
+    
+    config: dict or Callable
+        Dictionary of configuration options for the datastore. The following keys are required:
+        - root: The root path in the datastore where the data will be saved. (needs to be in the format expected by the datastore)
+            - example: 's3://bucket-name/path/to/root'
+            - example: 'gs://bucket-name/path/to/root'
+            - example: 'https://myblockacc.blob.core.windows.net/metaflow/'
+        - role_arn (optional): AWS IAM role to access s3 bucket (only when `type` is 's3')
+        - session_vars (optional): AWS session variables to access s3 bucket (only when `type` is 's3')
+        - client_params (optional): AWS client parameters to access s3 bucket (only when `type` is 's3')
     """
     ...