Skip to content

Craft specs

A craft spec is the same string you would pass to sktime.registry.craft: a class call, including constructor kwargs, with no imports. Pass it to Server as (id, spec), or to tserve as id=spec. Predict uses that name as model, not the spec string.

Catalog models cover published checkpoints. A spec is how you load a checkpoint, revision, or configuration the catalog does not name.

from tserve.server import Server

Server(
    model=[
        "chronos_bolt",
        (
            "ttm-local",
            'TinyTimeMixerForecaster(model_path="ibm-granite/granite-timeseries-ttm-r3", '
            'revision="52-16-dec-52-r3", fit_strategy="zero-shot")',
        ),
    ],
    host="127.0.0.1",
    port=8000,
).run()

{
  "models": [
    {"id": "naive", "executor": "sktime", "source": "registry"},
    {"id": "chronos_bolt", "executor": "sktime", "source": "registry"},
    {"id": "ttm-local", "executor": "sktime", "source": "craft"}
  ]
}

From the command line

A token with = is a craft spec, split on the first =. Catalog models have no =. Everything after the first = is the spec, so kwargs may contain = too. Quote the whole token so constructor quotes survive the shell:

tserve chronos_bolt \
  'ttm-local=TinyTimeMixerForecaster(model_path="ibm-granite/granite-timeseries-ttm-r3", revision="52-16-dec-52-r3", fit_strategy="zero-shot")'

Docker is the same argv after the image name. A token that looks like ClassName(...) with no id= is rejected; wrap it as id=spec. Flag details: CLI.

Rules

  • The spec must be a non-empty string that craft turns into a sktime BaseForecaster instance. A class name without parentheses ("NaiveForecaster") is rejected.
  • A bare spec in Python model is treated as an unknown registry model. Wrap it as (id, spec).
  • Models must be unique across the whole list. A collision — including with a registry model — raises ValueError before the second load.
  • Registry models, live objects, craft specs, and saved models mix freely in one model list.
  • The estimator's own dependencies have to be installed; TServe only adds the ones its extras declare.
  • The spec is evaluated only in your process at startup, never from an HTTP model field.

Each spec is loaded and warmed up like any other model, so startup pays the download and warmup cost once.