hsml.model_registry #

ModelRegistry #

project_name `property` #

Name of the project the registry is connected to.

project_path `property` #

Path of the project the registry is connected to.

project_id `property` #

Id of the project the registry is connected to.

shared_registry_project_name `property` #

Name of the project the shared model registry originates from.

model_registry_id `property` #

Id of the model registry.

tensorflow `property` #

Module for exporting a TensorFlow model.

sklearn `property` #

Module for exporting a sklearn model.

torch `property` #

Module for exporting a torch model.

python `property` #

Module for exporting a generic Python model.

llm `property` #

Module for exporting a Large Language Model.

get_model #

get_model(
    name: str, version: int | None = None
) -> model.Model | None

Get a model entity from the model registry.

Getting a model from the Model Registry means getting its metadata handle so you can subsequently download the model directory.

PARAMETER	DESCRIPTION
`name`	Name of the model to get. TYPE: `str`
`version`	Version of the model to retrieve, defaults to `None` and will return the `version=1`. TYPE: `int \| None` DEFAULT: `None`

RETURNS	DESCRIPTION
`model.Model \| None`	The model metadata object or `None` if it does not exist.

RAISES	DESCRIPTION
`hopsworks.client.exceptions.RestAPIError`	If unable to retrieve model from the model registry.

get_models #

get_models(name: str) -> list[model.Model]

Get all model entities from the model registry for a specified name.

Getting all models from the Model Registry for a given name returns a list of model entities, one for each version registered under the specified model name.

PARAMETER	DESCRIPTION
`name`	Name of the model to get. TYPE: `str`

RETURNS	DESCRIPTION
`list[model.Model]`	A list of model metadata objects.

RAISES	DESCRIPTION
`hopsworks.client.exceptions.RestAPIError`	If unable to retrieve model versions from the model registry.

get_best_model #

get_best_model(
    name: str, metric: str, direction: str
) -> model.Model | None

Get the best performing model entity from the model registry.

Getting the best performing model from the Model Registry means specifying in addition to the name, also a metric name corresponding to one of the keys in the training_metrics dict of the model and a direction. For example, to get the model version with the highest accuracy, specify metric='accuracy' and direction='max'.

PARAMETER	DESCRIPTION
`name`	Name of the model to get. TYPE: `str`
`metric`	Name of the key in the training metrics field to compare. TYPE: `str`
`direction`	'max' to get the model entity with the highest value of the set metric, or 'min' for the lowest. TYPE: `str`

RETURNS	DESCRIPTION
`model.Model \| None`	The model metadata object or `None` if it does not exist.

RAISES	DESCRIPTION
`hopsworks.client.exceptions.RestAPIError`	If unable to retrieve model from the model registry.

hf_download #

hf_download(
    hugging_face_model_id: str,
    hf_token: str | None = None,
    selected_formats: list[str] | None = None,
    selected_variants: list[str] | None = None,
    selected_filenames: list[str] | None = None,
    timeout: int = 36000,
    poll_interval: int = 5,
) -> model.Model

Import a model from HuggingFace into this Model Registry.

Kicks off the asynchronous server-side download, polls until the import reaches a terminal status, and returns the registered model on success.

import hopsworks

project = hopsworks.login()
mr = project.get_model_registry()

# Public model — minimal call.
m = mr.hf_download("Qwen/Qwen2.5-0.5B")

# Gated / private — supply an HF access token.
m = mr.hf_download("meta-llama/Llama-3.2-1B", hf_token="hf_xxx")

# Multi-quant repo — pick exactly one variant.
m = mr.hf_download(
    "unsloth/Kimi-K2.6-GGUF", selected_variants=["UD-Q4_K_XL"]
)

PARAMETER	DESCRIPTION
`hugging_face_model_id`	HF repo id (`owner/repo`) or full HuggingFace URL. TYPE: `str`
`hf_token`	Optional HuggingFace access token; required for gated/private repos. TYPE: `str \| None` DEFAULT: `None`
`selected_formats`	Weight-format keys to download (`safetensors`, `pytorch`, `tensorflow`, `flax`, `gguf`, `onnx`, `openvino`, `coreml`, `tflite`). When omitted, the backend picks one with the precedence safetensors > pytorch > tensorflow > flax > … so multi-format repos don't pull every copy of the same weights. TYPE: `list[str] \| None` DEFAULT: `None`
`selected_variants`	GGUF / unsloth quantisation tags to keep (`Q4_K_M`, `UD-Q4_K_XL`, `BF16`, …). Acts as a sub-filter — files with a quant tag are only downloaded if their tag is in the list. Files without a tag (root README, config, …) are unaffected. TYPE: `list[str] \| None` DEFAULT: `None`
`selected_filenames`	Explicit per-file allowlist. When non-empty this overrides `selected_formats` / `selected_variants` and the backend downloads exactly these paths. TYPE: `list[str] \| None` DEFAULT: `None`
`timeout`	Maximum seconds to wait for the import to reach a terminal state. TYPE: `int` DEFAULT: `36000`
`poll_interval`	Seconds between status polls. TYPE: `int` DEFAULT: `5`

RETURNS	DESCRIPTION
`model.Model`	The newly registered model entity.

RAISES	DESCRIPTION
`hopsworks.client.exceptions.HuggingFaceImportException`	If the backend fails the import with a known terminal error. The exception's `error_code` field carries the stable code (`auth_required`, `model_not_found`, `not_found_or_auth_required`, `no_disk_space`, `fetch_failed`, etc.).
`hopsworks.client.exceptions.RestAPIError`	For unexpected server errors that are not part of the HF import error set.

hsml.model_registry #

ModelRegistry #

project_name property #

project_path property #

project_id property #

shared_registry_project_name property #

model_registry_id property #

tensorflow property #

sklearn property #

torch property #

python property #

llm property #