flytekitplugins.kfpytorch.task

This Plugin adds the capability of running distributed pytorch training to Flyte using backend plugins, natively on Kubernetes. It leverages Pytorch Job <https://github.com/kubeflow/pytorch-operator>_ Plugin from kubeflow.

Directory

Classes

Methods

Variables

Methods

spawn_helper()

def spawn_helper(
    fn: bytes,
    raw_output_prefix: str,
    checkpoint_dest: str,
    checkpoint_src: str,
    kwargs,
) -> flytekitplugins.kfpytorch.task.ElasticWorkerResult

Help to spawn worker processes.

The purpose of this function is to 1) be pickleable so that it can be used with the multiprocessing start method spawn and 2) to call a cloudpickle-serialized function passed to it. This function itself doesn’t have to be pickleable. Without such a helper task functions, which are not pickleable, couldn’t be used with the start method spawn.

flytekitplugins.kfpytorch.task.Elastic

Configuration for torch elastic training <https://pytorch.org/docs/stable/elastic/run.html>_.

Use this to run single- or multi-node distributed pytorch elastic training on k8s.

Single-node elastic training is executed in a k8s pod when nnodes is set to 1. Multi-node training is executed otherwise using a Pytorch Job <https://github.com/kubeflow/training-operator>_.

Like torchrun, this plugin sets the environment variable OMP_NUM_THREADS to 1 if it is not set. Please see https://pytorch.org/tutorials/recipes/recipes/tuning_guide.html for potential performance improvements. To change OMP_NUM_THREADS, specify it in the environment dict of the flytekit task decorator or via pyflyte run --env.

class Elastic(
    nnodes: typing.Union[int, str],
    nproc_per_node: int,
    start_method: str,
    monitor_interval: int,
    max_restarts: int,
    rdzv_configs: typing.Dict[str, typing.Any],
    increase_shared_mem: bool,
    run_policy: typing.Optional[flytekitplugins.kfpytorch.task.RunPolicy],
)

flytekitplugins.kfpytorch.task.ElasticWorkerResult

A named tuple representing the result of a torch elastic worker process.

Attributes: return_value (Any): The value returned by the task function in the worker process. decks (list[flytekit.Deck]): A list of flytekit Deck objects created in the worker process.

flytekitplugins.kfpytorch.task.Master

Configuration for master replica group. Master should always have 1 replica, so we don’t need a replicas field

class Master(
    image: typing.Optional[str],
    requests: typing.Optional[flytekit.core.resources.Resources],
    limits: typing.Optional[flytekit.core.resources.Resources],
    restart_policy: typing.Optional[flytekitplugins.kfpytorch.task.RestartPolicy],
)

flytekitplugins.kfpytorch.task.PyTorch

Configuration for an executable PyTorch Job <https://github.com/kubeflow/pytorch-operator>_. Use this to run distributed PyTorch training on Kubernetes. Please notice, in most cases, you should not worry about the configuration of the master and worker groups. The default configuration should work. The only field you should change is the number of workers. Both replicas will use the same image, and the same resources inherited from task function decoration.

class PyTorch(
    master: flytekitplugins.kfpytorch.task.Master,
    worker: flytekitplugins.kfpytorch.task.Worker,
    run_policy: typing.Optional[flytekitplugins.kfpytorch.task.RunPolicy],
    num_workers: typing.Optional[int],
    increase_shared_mem: bool,
)

flytekitplugins.kfpytorch.task.PyTorchFunctionTask

Plugin that submits a PyTorchJob (see https://github.com/kubeflow/pytorch-operator) defined by the code within the _task_function to k8s cluster.

class PyTorchFunctionTask(
    task_config: flytekitplugins.kfpytorch.task.PyTorch,
    task_function: typing.Callable,
    kwargs,
)

Methods

compile()

def compile(
    ctx: flytekit.core.context_manager.FlyteContext,
    args,
    kwargs,
) -> typing.Union[typing.Tuple[flytekit.core.promise.Promise], flytekit.core.promise.Promise, flytekit.core.promise.VoidPromise, NoneType]

Generates a node that encapsulates this task in a workflow definition.

compile_into_workflow()

def compile_into_workflow(
    ctx: FlyteContext,
    task_function: Callable,
    kwargs,
) -> Union[_dynamic_job.DynamicJobSpec, _literal_models.LiteralMap]

In the case of dynamic workflows, this function will produce a workflow definition at execution time which will then proceed to be executed.

construct_node_metadata()

def construct_node_metadata()

Used when constructing the node that encapsulates this task as part of a broader workflow definition.

dispatch_execute()

def dispatch_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    input_literal_map: flytekit.models.literals.LiteralMap,
) -> typing.Union[flytekit.models.literals.LiteralMap, flytekit.models.dynamic_job.DynamicJobSpec, typing.Coroutine]

This method translates Flyte’s Type system based input values and invokes the actual call to the executor This method is also invoked during runtime.

• VoidPromise is returned in the case when the task itself declares no outputs.
• Literal Map is returned when the task returns either one more outputs in the declaration. Individual outputs may be none
• DynamicJobSpec is returned when a dynamic workflow is executed

dynamic_execute()

def dynamic_execute(
    task_function: Callable,
    kwargs,
) -> Any

By the time this function is invoked, the local_execute function should have unwrapped the Promises and Flyte literal wrappers so that the kwargs we are working with here are now Python native literal values. This function is also expected to return Python native literal values.

Since the user code within a dynamic task constitute a workflow, we have to first compile the workflow, and then execute that workflow.

When running for real in production, the task would stop after the compilation step, and then create a file representing that newly generated workflow, instead of executing it.

execute()

def execute(
    kwargs,
) -> Any

This method will be invoked to execute the task. If you do decide to override this method you must also handle dynamic tasks or you will no longer be able to use the task as a dynamic task generator.

find_lhs()

def find_lhs()

get_command()

def get_command(
    settings: SerializationSettings,
) -> List[str]

Returns the command which should be used in the container definition for the serialized version of this task registered on a hosted Flyte platform.

get_config()

def get_config(
    settings: SerializationSettings,
) -> Optional[Dict[str, str]]

Returns the task config as a serializable dictionary. This task config consists of metadata about the custom defined for this task.

get_container()

def get_container(
    settings: SerializationSettings,
) -> _task_model.Container

Returns the container definition (if any) that is used to run the task on hosted Flyte.

get_custom()

def get_custom(
    settings: flytekit.configuration.SerializationSettings,
) -> typing.Dict[str, typing.Any]

Return additional plugin-specific custom data (if any) as a serializable dictionary.

get_default_command()

def get_default_command(
    settings: SerializationSettings,
) -> List[str]

Returns the default pyflyte-execute command used to run this on hosted Flyte platforms.

get_extended_resources()

def get_extended_resources(
    settings: SerializationSettings,
) -> Optional[tasks_pb2.ExtendedResources]

Returns the extended resources to allocate to the task on hosted Flyte.

get_image()

def get_image(
    settings: SerializationSettings,
) -> str

Update image spec based on fast registration usage, and return string representing the image

get_input_types()

def get_input_types()

Returns the names and python types as a dictionary for the inputs of this task.

get_k8s_pod()

def get_k8s_pod(
    settings: SerializationSettings,
) -> _task_model.K8sPod

Returns the kubernetes pod definition (if any) that is used to run the task on hosted Flyte.

get_sql()

def get_sql(
    settings: flytekit.configuration.SerializationSettings,
) -> typing.Optional[flytekit.models.task.Sql]

Returns the Sql definition (if any) that is used to run the task on hosted Flyte.

get_type_for_input_var()

def get_type_for_input_var(
    k: str,
    v: typing.Any,
) -> typing.Type[typing.Any]

Returns the python type for an input variable by name.

get_type_for_output_var()

def get_type_for_output_var(
    k: str,
    v: typing.Any,
) -> typing.Type[typing.Any]

Returns the python type for the specified output variable by name.

local_execute()

def local_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    kwargs,
) -> typing.Union[typing.Tuple[flytekit.core.promise.Promise], flytekit.core.promise.Promise, flytekit.core.promise.VoidPromise, typing.Coroutine, NoneType]

This function is used only in the local execution path and is responsible for calling dispatch execute. Use this function when calling a task with native values (or Promises containing Flyte literals derived from Python native values).

local_execution_mode()

def local_execution_mode()

post_execute()

def post_execute(
    user_params: typing.Optional[flytekit.core.context_manager.ExecutionParameters],
    rval: typing.Any,
) -> typing.Any

Post execute is called after the execution has completed, with the user_params and can be used to clean-up, or alter the outputs to match the intended tasks outputs. If not overridden, then this function is a No-op

pre_execute()

def pre_execute(
    user_params: typing.Optional[flytekit.core.context_manager.ExecutionParameters],
) -> typing.Optional[flytekit.core.context_manager.ExecutionParameters]

This is the method that will be invoked directly before executing the task method and before all the inputs are converted. One particular case where this is useful is if the context is to be modified for the user process to get some user space parameters. This also ensures that things like SparkSession are already correctly setup before the type transformers are called

This should return either the same context of the mutated context

reset_command_fn()

def reset_command_fn()

Resets the command which should be used in the container definition of this task to the default arguments. This is useful when the command line is overridden at serialization time.

sandbox_execute()

def sandbox_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    input_literal_map: flytekit.models.literals.LiteralMap,
) -> flytekit.models.literals.LiteralMap

Call dispatch_execute, in the context of a local sandbox execution. Not invoked during runtime.

set_command_fn()

def set_command_fn(
    get_command_fn: Optional[Callable[[SerializationSettings], List[str]]],
)

By default, the task will run on the Flyte platform using the pyflyte-execute command. However, it can be useful to update the command with which the task is serialized for specific cases like running map tasks (“pyflyte-map-execute”) or for fast-executed tasks.

set_resolver()

def set_resolver(
    resolver: TaskResolverMixin,
)

By default, flytekit uses the DefaultTaskResolver to resolve the task. This method allows the user to set a custom task resolver. It can be useful to override the task resolver for specific cases like running tasks in the jupyter notebook.

Properties

flytekitplugins.kfpytorch.task.PytorchElasticFunctionTask

Plugin for distributed training with torch elastic/torchrun (see https://pytorch.org/docs/stable/elastic/run.html).

class PytorchElasticFunctionTask(
    task_config: flytekitplugins.kfpytorch.task.Elastic,
    task_function: typing.Callable,
    kwargs,
)

Methods

compile()

def compile(
    ctx: flytekit.core.context_manager.FlyteContext,
    args,
    kwargs,
) -> typing.Union[typing.Tuple[flytekit.core.promise.Promise], flytekit.core.promise.Promise, flytekit.core.promise.VoidPromise, NoneType]

Generates a node that encapsulates this task in a workflow definition.

compile_into_workflow()

def compile_into_workflow(
    ctx: FlyteContext,
    task_function: Callable,
    kwargs,
) -> Union[_dynamic_job.DynamicJobSpec, _literal_models.LiteralMap]

In the case of dynamic workflows, this function will produce a workflow definition at execution time which will then proceed to be executed.

construct_node_metadata()

def construct_node_metadata()

Used when constructing the node that encapsulates this task as part of a broader workflow definition.

dispatch_execute()

def dispatch_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    input_literal_map: flytekit.models.literals.LiteralMap,
) -> typing.Union[flytekit.models.literals.LiteralMap, flytekit.models.dynamic_job.DynamicJobSpec, typing.Coroutine]

This method translates Flyte’s Type system based input values and invokes the actual call to the executor This method is also invoked during runtime.

• VoidPromise is returned in the case when the task itself declares no outputs.
• Literal Map is returned when the task returns either one more outputs in the declaration. Individual outputs may be none
• DynamicJobSpec is returned when a dynamic workflow is executed

dynamic_execute()

def dynamic_execute(
    task_function: Callable,
    kwargs,
) -> Any

By the time this function is invoked, the local_execute function should have unwrapped the Promises and Flyte literal wrappers so that the kwargs we are working with here are now Python native literal values. This function is also expected to return Python native literal values.

Since the user code within a dynamic task constitute a workflow, we have to first compile the workflow, and then execute that workflow.

When running for real in production, the task would stop after the compilation step, and then create a file representing that newly generated workflow, instead of executing it.

execute()

def execute(
    kwargs,
) -> typing.Any

This method will be invoked to execute the task.

Handles the exception scope for the _execute method.

find_lhs()

def find_lhs()

get_command()

def get_command(
    settings: SerializationSettings,
) -> List[str]

Returns the command which should be used in the container definition for the serialized version of this task registered on a hosted Flyte platform.

get_config()

def get_config(
    settings: SerializationSettings,
) -> Optional[Dict[str, str]]

Returns the task config as a serializable dictionary. This task config consists of metadata about the custom defined for this task.

get_container()

def get_container(
    settings: SerializationSettings,
) -> _task_model.Container

Returns the container definition (if any) that is used to run the task on hosted Flyte.

get_custom()

def get_custom(
    settings: flytekit.configuration.SerializationSettings,
) -> typing.Optional[typing.Dict[str, typing.Any]]

Return additional plugin-specific custom data (if any) as a serializable dictionary.

get_default_command()

def get_default_command(
    settings: SerializationSettings,
) -> List[str]

Returns the default pyflyte-execute command used to run this on hosted Flyte platforms.

get_extended_resources()

def get_extended_resources(
    settings: SerializationSettings,
) -> Optional[tasks_pb2.ExtendedResources]

Returns the extended resources to allocate to the task on hosted Flyte.

get_image()

def get_image(
    settings: SerializationSettings,
) -> str

Update image spec based on fast registration usage, and return string representing the image

get_input_types()

def get_input_types()

Returns the names and python types as a dictionary for the inputs of this task.

get_k8s_pod()

def get_k8s_pod(
    settings: SerializationSettings,
) -> _task_model.K8sPod

Returns the kubernetes pod definition (if any) that is used to run the task on hosted Flyte.

get_sql()

def get_sql(
    settings: flytekit.configuration.SerializationSettings,
) -> typing.Optional[flytekit.models.task.Sql]

Returns the Sql definition (if any) that is used to run the task on hosted Flyte.

get_type_for_input_var()

def get_type_for_input_var(
    k: str,
    v: typing.Any,
) -> typing.Type[typing.Any]

Returns the python type for an input variable by name.

get_type_for_output_var()

def get_type_for_output_var(
    k: str,
    v: typing.Any,
) -> typing.Type[typing.Any]

Returns the python type for the specified output variable by name.

local_execute()

def local_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    kwargs,
) -> typing.Union[typing.Tuple[flytekit.core.promise.Promise], flytekit.core.promise.Promise, flytekit.core.promise.VoidPromise, typing.Coroutine, NoneType]

This function is used only in the local execution path and is responsible for calling dispatch execute. Use this function when calling a task with native values (or Promises containing Flyte literals derived from Python native values).

local_execution_mode()

def local_execution_mode()

post_execute()

def post_execute(
    user_params: typing.Optional[flytekit.core.context_manager.ExecutionParameters],
    rval: typing.Any,
) -> typing.Any

Post execute is called after the execution has completed, with the user_params and can be used to clean-up, or alter the outputs to match the intended tasks outputs. If not overridden, then this function is a No-op

pre_execute()

def pre_execute(
    user_params: typing.Optional[flytekit.core.context_manager.ExecutionParameters],
) -> typing.Optional[flytekit.core.context_manager.ExecutionParameters]

This is the method that will be invoked directly before executing the task method and before all the inputs are converted. One particular case where this is useful is if the context is to be modified for the user process to get some user space parameters. This also ensures that things like SparkSession are already correctly setup before the type transformers are called

This should return either the same context of the mutated context

reset_command_fn()

def reset_command_fn()

Resets the command which should be used in the container definition of this task to the default arguments. This is useful when the command line is overridden at serialization time.

sandbox_execute()

def sandbox_execute(
    ctx: flytekit.core.context_manager.FlyteContext,
    input_literal_map: flytekit.models.literals.LiteralMap,
) -> flytekit.models.literals.LiteralMap

Call dispatch_execute, in the context of a local sandbox execution. Not invoked during runtime.

set_command_fn()

def set_command_fn(
    get_command_fn: Optional[Callable[[SerializationSettings], List[str]]],
)

By default, the task will run on the Flyte platform using the pyflyte-execute command. However, it can be useful to update the command with which the task is serialized for specific cases like running map tasks (“pyflyte-map-execute”) or for fast-executed tasks.

set_resolver()

def set_resolver(
    resolver: TaskResolverMixin,
)

By default, flytekit uses the DefaultTaskResolver to resolve the task. This method allows the user to set a custom task resolver. It can be useful to override the task resolver for specific cases like running tasks in the jupyter notebook.

Properties

flytekitplugins.kfpytorch.task.RunPolicy

RunPolicy describes some policy to apply to the execution of a kubeflow job.

class RunPolicy(
    clean_pod_policy: <enum 'CleanPodPolicy'>,
    ttl_seconds_after_finished: typing.Optional[int],
    active_deadline_seconds: typing.Optional[int],
    backoff_limit: typing.Optional[int],
)

flytekitplugins.kfpytorch.task.Worker

class Worker(
    image: typing.Optional[str],
    requests: typing.Optional[flytekit.core.resources.Resources],
    limits: typing.Optional[flytekit.core.resources.Resources],
    replicas: typing.Optional[int],
    restart_policy: typing.Optional[flytekitplugins.kfpytorch.task.RestartPolicy],
)