kfp-toolbox 0.1.0

The toolbox for kfp (Kubeflow Pipelines SDK)

kfp-toolbox

kfp-toolbox is a Python library that provides useful tools for kfp (Kubeflow Pipelines SDK).

Installation

pip install kfp-toolbox

Usage

spec

from kfp_toolbox import spec

The spec decorator specifies the computing resources to be used by the component.

To apply this to a Python function-based component, it must be added outside of the component decorator.

@spec(cpu="2", memory="16G")
@dsl.component
def component_function():
    ...

For other components, wrap the component as a function.

component = kfp.components.load_component_from_file("path/to/component.yaml")
component = spec(cpu="2", memory="16G")(component)

If multiple spec decorators are stacked, the one placed further out will take precedence. For example, suppose you have created an alias default_spec. If you want to overwrite part of it, place a new spec decorator outside of the default_spec decorator to overwrite it.

default_spec = spec(cpu="2", memory="16G")

@spec(cpu="1")
@default_spec
@dsl.component
def component_function():
    ...

See all available options here:

option	type	description	examples
name	str	Display name	"Component NAME"
cpu	str	CPU limit	"1", "500m", ... ("m" means 1/1000)
memory	str	Memory limit	"512K", "16G", ...
gpu	str	GPU limit	"1", "2", ...
accelerator	str	Accelerator type	"NVIDIA_TESLA_K80", "TPU_V3", ...
caching	bool	Enable caching	True or False