Truss

0.15.12 · active · verified Mon Apr 13

Truss is a Python library that provides a seamless bridge from model development to model delivery. It allows data scientists to containerize, test, and deploy machine learning models as production-ready services with ease, often to platforms like Baseten. The current stable version is 0.15.12, with a fairly active release cadence, frequently releasing patch versions.

Warnings

breaking Starting with version 0.15.13, Truss will change its default Python version for new models to 3.13 and officially deprecate support for Python 3.9. Models currently using Python 3.9 will need to update their Python version.
Fix: For models deployed with Truss 0.15.13 or newer, ensure your `config.yaml` explicitly specifies a supported Python version (e.g., `python_version: '3.10'`) or upgrade your model's environment to Python 3.10 or newer.
gotcha Truss utilizes certain environment variables internally for its operations and configuration. Overriding these with your own custom environment variables can lead to unexpected behavior, build failures, or runtime errors within your deployed model.
Fix: Consult the official Truss documentation for a list of reserved environment variables to avoid collisions. Consider prefixing your custom model-specific environment variables (e.g., `MY_MODEL_VAR_NAME`) to minimize the risk of conflict.
gotcha Since version 0.15.10, the `config.yaml` file for Truss models adheres to a JSON schema, with validation occurring during build and push operations. This can cause older `config.yaml` files with deprecated or invalid syntax to fail validation.
Fix: If you encounter validation errors, refer to the latest `config.yaml` schema documentation or try initializing a new Truss project and migrating your model code to ensure your configuration is up-to-date and compliant.

Install

pip install truss Install stable version

Imports

truss
```
import truss
```

Quickstart

This quickstart demonstrates how to initialize a new Truss project and add a basic `predict` function. The example includes simulating model data loading via an environment variable. To actually serve and test the model, navigate into the created directory (`cd my_truss_model`) and run `truss serve` from your terminal. For deployment to platforms like Baseten, an API key (e.g., `os.environ.get('BASETEN_API_KEY', '')`) would typically be required for `truss.push()` operations.

import truss
import os
import pathlib
import shutil

# Define model name and directory
model_name = "my_truss_model"
model_dir = pathlib.Path(f"./{model_name}")

# Clean up previous runs if any
if model_dir.exists():
    shutil.rmtree(model_dir)

# Initialize a new Truss model structure, creating the directory
truss.init(model_name)

# Modify the model.py file to add a simple predict function
model_py_path = model_dir / "model" / "model.py"
with open(model_py_path, "w") as f:
    f.write("""
import os

class Model:
    def __init__(self):
        self._data = None

    def load(self):
        # Simulate loading a model artifact or configuration
        self._data = os.environ.get("TRUSS_MODEL_CONFIG", "default_config_value")

    def predict(self, model_input):
        return f"Hello, {model_input}! Model loaded with config: {self._data}"
""")

print(f"Truss model initialized in: {model_dir.resolve()}")
print(f"To serve locally: Change directory to '{model_name}' (cd {model_name}) and run 'truss serve' in your terminal.")

# To clean up the created model directory (uncomment to run):
# shutil.rmtree(model_dir)

view raw JSON →