Components

Components are the configurable building blocks of the model that an engine assembles for a run. They are selected under the components section of the engine configuration, one per kind:

  • model, the GNN itself,

  • transformer, which applies per-feature value transforms such as a logarithmic or inverse-hyperbolic-sine mapping,

  • normalizer, which rescales feature and target tensors into a conditioned range.

  • optimizer, the optimization algorithm that updates model weights, and

  • loss, the objective minimized during training.

A component is the composition of two objects: it is a PyTorch Module and an IceGraph plugin at once. From the module it inherits parameters, buffers, device placement, and checkpoint serialization; from the plugin it inherits the identity, configuration, and lifecycle model.

A component tracks the device it currently occupies. On attach, a component may be checked against any contract issued by the policy.