Functional

TorchLean functional helpers in the style of torch.* building blocks.

These are derived operations built from Runtime.Autograd.Torch.Ops, so eager execution and typed graph construction share the same model and loss definitions.

For background, see the PyTorch documentation for torch.nn.functional, torch.autograd, and checkpointing, together with the standard reverse-mode AD references by Linnainmaa and by Griewank and Walther.

Functional.EinsumDynamic is deliberately not re-exported here. It holds the runtime-checked, string-driven einsum?, which is by far the most expensive declaration in this subtree and which nothing in the runtime or the API layer calls. Runtime.Autograd.Model re-exports it, so it stays part of the public surface without sitting in front of the layer and trainer modules that every build has to get through.