HuggingFace runtime for MLServer

This package provides a MLServer runtime compatible with HuggingFace Transformers.


You can install the runtime, alongside mlserver, as:

pip install mlserver mlserver-huggingface

For further information on how to use MLServer with HuggingFace, you can check out this worked out example.

Content Types

The HuggingFace runtime will always decode the input request using its own built-in codec. Therefore, content type annotations at the request level will be ignored. Note that this doesn’t include input-level content type annotations, which will be respected as usual.


The HuggingFace runtime exposes a couple extra parameters which can be used to customise how the runtime behaves. These settings can be added under the parameters.extra section of your model-settings.json file, e.g.

  "name": "qa",
  "implementation": "mlserver_huggingface.HuggingFaceRuntime",
  "parameters": {
    "extra": {
      "task": "question-answering",
      "optimum_model": true


These settings can also be injected through environment variables prefixed with MLSERVER_MODEL_HUGGINGFACE_, e.g.


Loading models

Local models

It is possible to load a local model into a HuggingFace pipeline by specifying the model artefact folder path in parameters.uri in model-settings.json.

HuggingFace models

Models in the HuggingFace hub can be loaded by specifying their name in parameters.extra.pretrained_model in model-settings.json.


If parameters.extra.pretrained_model is specified, it takes precedence over parameters.uri.


You can find the full reference of the accepted extra settings for the HuggingFace runtime below:

pydantic settings mlserver_huggingface.settings.HuggingFaceSettings

Parameters that apply only to HuggingFace models

  • env_prefix: str = MLSERVER_MODEL_HUGGINGFACE_

  • extra: Extra = ignore

field device: int = -1

Device in which this pipeline will be loaded (e.g., “cpu”, “cuda:1”, “mps”, or a GPU ordinal rank like 1).

field framework: str | None = None

The framework to use, either “pt” for PyTorch or “tf” for TensorFlow.

field inter_op_threads: int | None = None

Threads used for parallelism between independent operations. PyTorch: Tensorflow:

field intra_op_threads: int | None = None

Threads used within an individual op for parallelism. PyTorch: Tensorflow:

field model_kwargs: dict | None = None

model kwargs that should be loaded in the pipeline.

field optimum_model: bool = False

Flag to decide whether the pipeline should use a Optimum-optimised model or the standard Transformers model. Under the hood, this will enable the model to use the optimised ONNX runtime.

field pretrained_model: str | None = None

Name of the model that should be loaded in the pipeline.

field pretrained_tokenizer: str | None = None

Name of the tokenizer that should be loaded in the pipeline.

field task: str = ''

Pipeline task to load. You can see the available Optimum and Transformers tasks available in the links below:

field task_suffix: str = ''

Suffix to append to the base task name. Useful for, e.g. translation tasks which require a suffix on the task name to specify source and target.

property task_name