Run Inference on servers

Inference is the process of using a trained model to make predictions on new data. As this process can be compute-intensive, running on a dedicated server can be an interesting option. The huggingface_hub library provides an easy way to call a service that runs inference for hosted models. There are several services you can connect to:

• Inference API: a service that allows you to run accelerated inference on Model Database’s infrastructure for free. This service is a fast way to get started, test different models, and prototype AI products.
• Inference Endpoints: a product to easily deploy models to production. Inference is run by Model Database in a dedicated, fully managed infrastructure on a cloud provider of your choice.

These services can be called with the InferenceClient object. It acts as a replacement for the legacy InferenceApi client, adding specific support for tasks and handling inference on both Inference API and Inference Endpoints. Learn how to migrate to the new client in the Legacy InferenceAPI client section.

Getting started

Let’s get started with a text-to-image task:

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()

>>> image = client.text_to_image("An astronaut riding a horse on the moon.")
>>> image.save("astronaut.png")

We initialized an InferenceClient with the default parameters. The only thing you need to know is the task you want to perform. By default, the client will connect to the Inference API and select a model to complete the task. In our example, we generated an image from a text prompt. The returned value is a PIL.Image object that can be saved to a file.

Using a specific model

What if you want to use a specific model? You can specify it either as a parameter or directly at an instance level:

>>> from huggingface_hub import InferenceClient
# Initialize client for a specific model
>>> client = InferenceClient(model="prompthero/openjourney-v4")
>>> client.text_to_image(...)
# Or use a generic client but pass your model as an argument
>>> client = InferenceClient()
>>> client.text_to_image(..., model="prompthero/openjourney-v4")

Using a specific URL

The examples we saw above use the free-hosted Inference API. This proves to be very useful for prototyping and testing things quickly. Once you’re ready to deploy your model to production, you’ll need to use a dedicated infrastructure. That’s where Inference Endpoints comes into play. It allows you to deploy any model and expose it as a private API. Once deployed, you’ll get a URL that you can connect to using exactly the same code as before, changing only the model parameter:

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient(model="https://uu149rez6gw9ehej.eu-west-1.aws.endpoints.huggingface.cloud/deepfloyd-if")
# or
>>> client = InferenceClient()
>>> client.text_to_image(..., model="https://uu149rez6gw9ehej.eu-west-1.aws.endpoints.huggingface.cloud/deepfloyd-if")

Authentication

Calls made with the InferenceClient can be authenticated using a User Access Token. By default, it will use the token saved on your machine if you are logged in (check out how to login). If you are not logged in, you can pass your token as an instance parameter:

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient(token="hf_***")

Supported tasks

InferenceClient’s goal is to provide the easiest interface to run inference on Model Database models. It has a simple API that supports the most common tasks. Here is a list of the currently supported tasks:

Custom requests

However, it is not always possible to cover all use cases. For custom requests, the InferenceClient.post() method gives you the flexibility to send any request to the Inference API. For example, you can specify how to parse the inputs and outputs. In the example below, the generated image is returned as raw bytes instead of parsing it as a PIL Image. This can be helpful if you don’t have Pillow installed in your setup and just care about the binary content of the image. InferenceClient.post() is also useful to handle tasks that are not yet officially supported.

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> response = client.post(json={"inputs": "An astronaut riding a horse on the moon."}, model="stabilityai/stable-diffusion-2-1")
>>> response.content # raw bytes
b'...'

Async client

An async version of the client is also provided, based on asyncio and aiohttp. You can either install aiohttp directly or use the [inference] extra:

pip install --upgrade huggingface_hub[inference]
# or
# pip install aiohttp

After installation all async API endpoints are available via AsyncInferenceClient. Its initialization and APIs are strictly the same as the sync-only version.

# Code must be run in a asyncio concurrent context.
# $ python -m asyncio
>>> from huggingface_hub import AsyncInferenceClient
>>> client = AsyncInferenceClient()

>>> image = await client.text_to_image("An astronaut riding a horse on the moon.")
>>> image.save("astronaut.png")

>>> async for token in await client.text_generation("The Huggingface Hub is", stream=True):
...     print(token, end="")
 a platform for sharing and discussing ML-related content.

For more information about the asyncio module, please refer to the official documentation.

Advanced tips

In the above section, we saw the main aspects of InferenceClient. Let’s dive into some more advanced tips.

Timeout

When doing inference, there are two main causes for a timeout:

• The inference process takes a long time to complete.
• The model is not available, for example when Inference API is loading it for the first time.

InferenceClient has a global timeout parameter to handle those two aspects. By default, it is set to None, meaning that the client will wait indefinitely for the inference to complete. If you want more control in your workflow, you can set it to a specific value in seconds. If the timeout delay expires, an InferenceTimeoutError is raised. You can catch it and handle it in your code:

>>> from huggingface_hub import InferenceClient, InferenceTimeoutError
>>> client = InferenceClient(timeout=30)
>>> try:
...     client.text_to_image(...)
... except InferenceTimeoutError:
...     print("Inference timed out after 30s.")

Binary inputs

Some tasks require binary inputs, for example, when dealing with images or audio files. In this case, InferenceClient tries to be as permissive as possible and accept different types:

• raw bytes
• a file-like object, opened as binary (with open("audio.flac", "rb") as f: ...)
• a path (str or Path) pointing to a local file
• a URL (str) pointing to a remote file (e.g. https://...). In this case, the file will be downloaded locally before sending it to the Inference API.

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> client.image_classification("https://upload.wikimedia.org/wikipedia/commons/thumb/4/43/Cute_dog.jpg/320px-Cute_dog.jpg")
[{'score': 0.9779096841812134, 'label': 'Blenheim spaniel'}, ...]

Legacy InferenceAPI client

InferenceClient acts as a replacement for the legacy InferenceApi client. It adds specific support for tasks and handles inference on both Inference API and Inference Endpoints.

Here is a short guide to help you migrate from InferenceApi to InferenceClient.

Initialization

Change from

>>> from huggingface_hub import InferenceApi
>>> inference = InferenceApi(repo_id="bert-base-uncased", token=API_TOKEN)

to

>>> from huggingface_hub import InferenceClient
>>> inference = InferenceClient(model="bert-base-uncased", token=API_TOKEN)

Run on a specific task

Change from

>>> from huggingface_hub import InferenceApi
>>> inference = InferenceApi(repo_id="paraphrase-xlm-r-multilingual-v1", task="feature-extraction")
>>> inference(...)

to

>>> from huggingface_hub import InferenceClient
>>> inference = InferenceClient()
>>> inference.feature_extraction(..., model="paraphrase-xlm-r-multilingual-v1")

Run custom request

Change from

>>> from huggingface_hub import InferenceApi
>>> inference = InferenceApi(repo_id="bert-base-uncased")
>>> inference(inputs="The goal of life is [MASK].")
[{'sequence': 'the goal of life is life.', 'score': 0.10933292657136917, 'token': 2166, 'token_str': 'life'}]

to

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> response = client.post(json={"inputs": "The goal of life is [MASK]."}, model="bert-base-uncased")
>>> response.json()
[{'sequence': 'the goal of life is life.', 'score': 0.10933292657136917, 'token': 2166, 'token_str': 'life'}]

Run with parameters

Change from

>>> from huggingface_hub import InferenceApi
>>> inference = InferenceApi(repo_id="typeform/distilbert-base-uncased-mnli")
>>> inputs = "Hi, I recently bought a device from your company but it is not working as advertised and I would like to get reimbursed!"
>>> params = {"candidate_labels":["refund", "legal", "faq"]}
>>> inference(inputs, params)
{'sequence': 'Hi, I recently bought a device from your company but it is not working as advertised and I would like to get reimbursed!', 'labels': ['refund', 'faq', 'legal'], 'scores': [0.9378499388694763, 0.04914155602455139, 0.013008488342165947]}

to

>>> from huggingface_hub import InferenceClient
>>> client = InferenceClient()
>>> inputs = "Hi, I recently bought a device from your company but it is not working as advertised and I would like to get reimbursed!"
>>> params = {"candidate_labels":["refund", "legal", "faq"]}
>>> response = client.post(json={"inputs": inputs, "parameters": params}, model="typeform/distilbert-base-uncased-mnli")
>>> response.json()
{'sequence': 'Hi, I recently bought a device from your company but it is not working as advertised and I would like to get reimbursed!', 'labels': ['refund', 'faq', 'legal'], 'scores': [0.9378499388694763, 0.04914155602455139, 0.013008488342165947]}