Recently, ML practitioners have applied graph-based machine learning models for use cases including product recommendations and synthetic biology. Graphs, which are more structured than tensors, allow users to represent prior knowledge when developing learning algorithms. However, without any easy-to-use frameworks, building ML models with graph-based data structures is really hard. To make it easier for researchers and ML engineers to train graph-based ML models, Facebook has released GTM, an OSS framework for automatic differentiation with weighted finite-state transducers (WFSTs). WFSTs are used to combine different information sources and can be applied to train different types of models (e.g. an acoustic model and a language model) together. GTN provides tools to define, visualize, and perform operations on WFSTs.