TorchLean

TorchLean is the runtime front-end for training and execution.

This module is the user-facing wrapper around the lower-level runtime session implementation:

• write a model/loss once over a small Ops interface,
• choose backend := .eager (dynamic tape) or backend := .compiled (typed SSA/DAG),
• run forward, backward, and step with the same call shape.

Runtime.Autograd.TorchLean is the stable runtime namespace re-exported by NN.API.Runtime. Runtime.Autograd.Torch remains available as the lower-level session layer used internally by TorchLean and by linked compiled sessions.

This umbrella deliberately does not own model catalogs or RL objectives. Reusable architecture specifications live under NN.GraphSpec.Models.TorchLean, while differentiable PPO / actor-critic loss helpers live under NN.Runtime.RL.PolicyGradient.Autograd. Keeping those out of the runtime core makes the dependency graph easier to audit: this folder should provide tensors, ops, modules, sessions, losses, optim/training glue, and executable autodiff utilities.

Training helpers

Optimizers

Module wrappers (PyTorch-style)