torchx.pipelines

These modules are adapters used to run TorchX components as part of a pipeline to allow for more complex behaviors as well as for continuous deployment.

The runner and schedulers are designed to launch a single component quickly where as these adapters transform the component into something understandable by the specific pipeline provider so you can assemble a full pipeline with them.

torchx.pipelines.kfp

This module contains adapters for converting TorchX components into KubeFlow Pipeline components.

The current KFP adapters only support single node (1 role and 1 replica) components.

torchx.pipelines.kfp.adapter.component_from_app(app: torchx.specs.api.AppDef) → torchx.pipelines.kfp.adapter.ContainerFactory

component_from_app takes in a TorchX component/AppDef and returns a KFP ContainerOp factory. This is equivalent to the kfp.components.load_component_from_* methods.

>>> from torchx import specs
>>> from torchx.pipelines.kfp.adapter import component_from_app
>>> app_def = specs.AppDef(
...     name="trainer",
...     roles=[specs.Role("trainer", image="foo:latest")],
... )
>>> component_from_app(app_def)
<function component_from_app...>

torchx.pipelines.kfp.adapter.component_spec_from_app(app: torchx.specs.api.AppDef) → Tuple[str, torchx.specs.api.Role]

component_spec_from_app takes in a TorchX component and generates the yaml spec for it. Notably this doesn’t apply resources or port_maps since those must be applied at runtime which is why it returns the role spec as well.

>>> from torchx import specs
>>> from torchx.pipelines.kfp.adapter import component_spec_from_app
>>> app_def = specs.AppDef(
...     name="trainer",
...     roles=[specs.Role("trainer", image="foo:latest")],
... )
>>> component_spec_from_app(app_def)
('description: ...', Role(...))

class torchx.pipelines.kfp.adapter.ContainerFactory(*args, **kwargs)

ContainerFactory is a protocol that represents a function that when called produces a kfp.dsl.ContainerOp.