Custom models

The public RTJModel is an ordinary torch.nn.Module whose state-dict names match published RT-J checkpoints. Advanced users can replace decoder heads, freeze blocks, register hooks, compile the model, or train from a new initialization.

Use the meta backend when inspecting or transforming a large architecture without allocating weights. Custom input encoders must preserve the checkpoint contract or be trained jointly with the backbone.