register_model_provider

inference_models.developer_tools.register_model_provider

register_model_provider(provider_name, provider_handler)

Register a custom weights provider for model retrieval.

Allows registration of custom model providers that can be used with AutoModel.from_pretrained() and get_model_from_provider(). This enables loading models from custom sources beyond the default Roboflow provider.

Parameters:

• provider_name

  (str) –

  Unique name for the provider. Will be used as the weights_provider parameter in AutoModel.from_pretrained().

• provider_handler

  (WeightsProvider) –

  Callable that implements the provider interface. Must accept (model_id: str, api_key: Optional[str], **kwargs) and return a ModelMetadata object.

Examples:

Register a custom provider:

>>> from inference_models.developer_tools import (
...     register_model_provider,
...     ModelMetadata,
...     ModelPackageMetadata,
...     ONNXPackageDetails
... )
>>> from inference_models import BackendType, Quantization
>>>
>>> def my_custom_provider(model_id: str, api_key: str = None, **kwargs):
...     # Fetch model metadata from your custom source
...     return ModelMetadata(
...         model_id=model_id,
...         model_packages=[
...             ModelPackageMetadata(
...                 backend_type=BackendType.ONNX,
...                 quantization=Quantization.FP32,
...                 package_details=ONNXPackageDetails(
...                     download_url="https://my-server.com/model.onnx",
...                     md5_hash="abc123...",
...                     # ... other details
...                 ),
...                 # ... other metadata
...             )
...         ],
...         dependencies=[]
...     )
>>>
>>> # Register the provider
>>> register_model_provider("my_provider", my_custom_provider)
>>>
>>> # Now use it with AutoModel
>>> from inference_models import AutoModel
>>> model = AutoModel.from_pretrained(
...     "my-model-id",
...     weights_provider="my_provider"
... )

Register a provider for local file system:

>>> def local_file_provider(model_id: str, api_key: str = None, **kwargs):
...     base_path = kwargs.get("base_path", "/models")
...     model_path = f"{base_path}/{model_id}"
...
...     # Return metadata pointing to local files
...     return ModelMetadata(
...         model_id=model_id,
...         model_packages=[...],  # Configure packages
...         dependencies=[]
...     )
>>>
>>> register_model_provider("local", local_file_provider)
>>>
>>> model = AutoModel.from_pretrained(
...     "yolov8n",
...     weights_provider="local",
...     base_path="/my/models"
... )

Note

• Provider handlers must return a ModelMetadata object
• The provider name must be unique (will override existing providers)
• Provider handlers should handle authentication and error cases

See Also

• get_model_from_provider(): Retrieve models using registered providers
• ModelMetadata: Structure for model metadata
• ModelPackageMetadata: Structure for package metadata