Task outputs as artifacts

The most common way to create an artifact is to register a task output. Two things are required: the task must be declared with produces_artifacts=True, and the return value must be wrapped with flyte.artifacts.new().

import flyte
from flyte import artifacts
from flyte.io import File

env = flyte.TaskEnvironment(name="training")


@env.task(produces_artifacts=True)
async def train() -> File:
    weights = await File.from_local("model.pt")
    return artifacts.new(
        weights,
        artifacts.Metadata(
            name="trained-model",
            description="Classifier trained on the latest training set",
            attrs={"framework": "pytorch"},
        ),
    )

Every run of this task registers a new version of trained-model. Union records the producing run on the artifact, so you can always trace a version back to the execution that created it.

What can be an artifact

An artifact must be an offloaded value: a flyte.io.File, a flyte.io.Dir, or a flyte.io.DataFrame. Primitives, dataclasses, and Pydantic models cannot be artifacts. Wrap a raw dataframe with DataFrame.from_df() first.

The wrapped value must also be a top-level task output. Nesting an artifact inside a list, a dictionary, or a model raises an error. In a multi-output task, only the wrapped slot becomes an artifact:

@env.task(produces_artifacts=True)
async def train() -> tuple[File, float]:
    ...
    return artifacts.new(weights, metadata), accuracy

Metadata

flyte.artifacts.Metadata carries the identity of the artifact. Only name is required. If you leave out version, the version comes from the producing run, so every execution registers a distinct version. You can also set a description, string-valued attrs for searching and filtering, a kind ("model", "data", or "generic"), and a card.

For models there is a helper that fills in the standard fields and records the artifact’s kind as model:

card = artifacts.Card.create_from(content=card_html, format="html", card_type="model")

metadata = artifacts.Metadata.create_model_metadata(
    name="trained-model",
    description="A toy classifier",
    framework="PyTorch",
    architecture="ResNet50",
    task="Image Classification",
    serial_format="pt",
    card=card,
)

A card is a document attached to the artifact and rendered in the UI. flyte.artifacts.Card.create_from() accepts inline content or a local file, in formats including HTML, Markdown, JSON, YAML, CSV, and PNG.

Consuming artifacts in tasks

Downstream tasks take artifacts as ordinary typed inputs. Fetch a reference with flyte.remote.Artifact.get() and pass it to flyte.run():

from flyte.remote import Artifact

model = Artifact.get("trained-model")               # latest version
data = Artifact.get("training-set", version="v3")   # pinned version

run = flyte.run(evaluate, model=model, data=data)

The task body receives a plain File, Dir, or DataFrame and needs no artifact-specific code.