Hugging Face Hub

Integrate Hugging Face Hub with W&B Weave to track and analyze your machine learning applications

All code samples shown on this page are in Python.

This page explains how to integrate Hugging Face Hub with W&B Weave to track and analyze your machine learning applications. You’ll learn how to log model inferences, monitor function calls, and organize experiments using Weave’s tracing and versioning capabilities. By following the examples provided, you can capture insights, debug your applications, and compare different model configurations, all within the Weave web interface.

Try Hugging Face Hub with Weave in Google Colab Do you want to experiment with Hugging Face Hub and Weave without any of the setup? You can try the code samples shown here as a Jupyter Notebook on Google Colab.

Overview

Hugging Face Hub is a machine learning platform for creators and collaborators, offering a large collection of pre-trained models and datasets for many types of projects.

The huggingface_hub Python library provides a unified interface to run inference across multiple services for models hosted on the Hub. You can invoke these models using the InferenceClient.

Weave automatically captures traces for InferenceClient. To start tracking, call weave.init() and use the library as usual.

Prerequisites

Before you start, complete the following setup steps to install the required libraries and configure access to the Hugging Face Hub:

1. Before you can use huggingface_hub with Weave, you must install the necessary libraries, or upgrade to the latest versions. The following command installs or upgrades huggingface_hub and weave to the latest version if they’re already installed, and reduces installation output.

   pip install -U huggingface_hub weave -qqq
   

2. To use inference with a model on the Hugging Face Hub, set your User Access Token. You can either set the token from your Hugging Face Hub Settings page or programmatically. The following code sample prompts you to enter your HUGGINGFACE_TOKEN and sets the token as an environment variable.

   import os
   import getpass
   
   os.environ["HUGGINGFACE_TOKEN"] = getpass.getpass("Enter your Hugging Face Hub Token: ")
   

Basic tracing

Storing traces of language model applications in a central location is essential during development and production. These traces help with debugging and serve as datasets for improving your application.

Weave automatically captures traces for the InferenceClient. To start tracking, initialize Weave by calling weave.init(), then use the library as usual.

The following example demonstrates how to log inference calls to the Hugging Face Hub using Weave:

import weave
from huggingface_hub import InferenceClient

# Initialize Weave
weave.init(project_name="quickstart-huggingface")

# Initialize Hugging Face Inference Client
huggingface_client = InferenceClient(
    api_key=os.environ.get("HUGGINGFACE_TOKEN")
)

# Make a chat completion inference call to the Hugging Face Hub with the Llama-3.2-11B-Vision-Instruct model
image_url = "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
response = huggingface_client.chat_completion(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "image_url", "image_url": {"url": image_url}},
                {"type": "text", "text": "Describe this image in one sentence."},
            ],
        }
    ],
    max_tokens=500,
    seed=42,
)

After the previous code runs, Weave tracks and logs all LLM calls made with the Hugging Face Inference Client. You can view these traces in the Weave web interface.

Weave logs each inference call, providing details about inputs, outputs, and metadata.

Weave also renders the call as a chat view in the UI, displaying the entire chat history with the model.

Trace a function

To gain deeper insights into how data flows through your application, you can use @weave.op to track function calls. This captures inputs, outputs, and execution logic, helping with debugging and performance analysis.

By nesting multiple ops, you can build a structured tree of tracked functions. Weave also automatically versions your code, preserving intermediate states as you experiment, even before you commit changes to Git.

To start tracking, decorate the functions that you want to track with @weave.op.

In the following example, Weave tracks three functions: generate_image, check_image_correctness, and generate_image_and_check_correctness. These functions generate an image and validate whether it matches a given prompt.

import base64
from PIL import Image


def encode_image(pil_image):
    import io
    buffer = io.BytesIO()
    pil_image.save(buffer, format="JPEG")
    buffer.seek(0)
    encoded_image = base64.b64encode(buffer.read()).decode("utf-8")
    return f"data:image/jpeg;base64,{encoded_image}"


@weave.op
def generate_image(prompt: str):
    return huggingface_client.text_to_image(
        prompt=prompt,
        model="black-forest-labs/FLUX.1-schnell",
        num_inference_steps=4,
    )


@weave.op
def check_image_correctness(image: Image.Image, image_generation_prompt: str):
    return huggingface_client.chat_completion(
        model="meta-llama/Llama-3.2-11B-Vision-Instruct",
        messages=[
            {
                "role": "user",
                "content": [
                    {"type": "image_url", "image_url": {"url": encode_image(image)}},
                    {
                        "type": "text",
                        "text": f"Is this image correct for the prompt: {image_generation_prompt}? Answer with only one word: yes or no",
                    },
                ],
            }
        ],
        max_tokens=500,
        seed=42,
    ).choices[0].message.content


@weave.op
def generate_image_and_check_correctness(prompt: str):
    image = generate_image(prompt)
    return {
        "image": image,
        "is_correct": check_image_correctness(image, prompt),
    }


response = generate_image_and_check_correctness("A cute puppy")

Weave now logs all function calls wrapped with @weave.op, letting you analyze execution details in the Weave UI.

Weave also captures and visualizes function execution, helping you to understand data flow and logic within your application.

Use Models for experimentation

Managing LLM experiments can be challenging when multiple components are involved. The Weave Model class helps capture and organize experimental details, such as system prompts and model configurations, letting you compare different iterations.

In addition to versioning code and capturing inputs and outputs, a Model stores structured parameters that control application behavior. This helps you track which configurations produced the best results. You can also integrate a Weave Model with Weave Serve and Evaluations for further insights.

The following example defines a CityVisitRecommender model for travel recommendations. Each modification to its parameters generates a new version, supporting iterative experimentation.

import rich


class CityVisitRecommender(weave.Model):
    model: str
    temperature: float = 0.7
    max_tokens: int = 500
    seed: int = 42

    @weave.op()
    def predict(self, city: str) -> str:
        return huggingface_client.chat_completion(
            model=self.model,
            messages=[
                {
                    "role": "system",
                    "content": "You are a helpful assistant meant to suggest places to visit in a city",
                },
                {"role": "user", "content": city},
            ],
            max_tokens=self.max_tokens,
            temperature=self.temperature,
            seed=self.seed,
        ).choices[0].message.content


city_visit_recommender = CityVisitRecommender(
    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
    temperature=0.7,
    max_tokens=500,
    seed=42,
)
rich.print(city_visit_recommender.predict("New York City"))
rich.print(city_visit_recommender.predict("Paris"))

Weave automatically logs models and tracks different versions so you can analyze performance and experiment history.