In this example, we go through building your first Truss model. We’ll be using the HuggingFace transformers library to build a text classification model that can detect sentiment of text.

​

Step 1: Implementing the model

Set up imports for this model. In this example, we simply use the HuggingFace transformers library.

from transformers import pipeline

Every Truss model must implement a Model class. This class must have:

• an __init__ function
• a load function
• a predict function

In the __init__ function, set up any variables that will be used in the load and predict functions.

class Model:
    def __init__(self, **kwargs):
        self._model = None

In the load function of the Truss, we implement logic involved in downloading the model and loading it into memory. For this Truss example, we define a HuggingFace pipeline, and choose the text-classification task, which uses BERT for text classification under the hood.

Note that the load function runs once when the model starts.

    def load(self):
        self._model = pipeline("text-classification")

In the predict function of the Truss, we implement logic related to actual inference. For this example, we just call the HuggingFace pipeline that we set up in the load function.

    def predict(self, model_input):
        return self._model(model_input)

​

Step 2: Writing the config.yaml

Each Truss has a config.yaml file where we can configure options related to the deployment. It’s in this file where we can define requirements, resources, and runtime options like secrets and environment variables

​

Basic Options

In this section, we can define basic metadata about the model, such as the name, and the Python version to build with.

model_name: bert
python_version: py310
model_metadata:
  example_model_input: { "text": "Hello my name is {MASK}" }

​

Set up python requirements

In this section, we define any pip requirements that we need to run the model. To run this, we need PyTorch and Tranformers.

requirements:
  - torch==2.0.1
  - transformers==4.33.2

​

Configure the resources needed

In this section, we can configure resources needed to deploy this model. Here, we have no need for a GPU so we leave the accelerator section blank.

resources:
  accelerator: null
  cpu: '1'
  memory: 2Gi
  use_gpu: false

​

Other config options

Truss also has provisions for adding other runtime options packages. In this example, we don’t need these, so we leave this empty for now.

secrets: {}
system_packages: []
environment_variables: {}
external_package_dirs: []

​

Step 3: Deploying & running inference

Deploy the model with the following command:

$ truss push

And then you can performance inference with:

$ truss predict -d '"Truss is awesome!"'

from transformers import pipeline


class Model:
    def __init__(self, **kwargs):
        self._model = None

    def load(self):
        self._model = pipeline("text-classification")

    def predict(self, model_input):
        return self._model(model_input)