Torch Core

Torch-style runtime front-end for eager execution, typed graphs, and training helpers.

• Core.Types: public handles, options, and parameter wrappers.
• Core.Session: eager session state, CUDA bridge, and tape lifecycle helpers.
• Core.Ops: eager tensor operations.
• Core.BackwardOptim: eager backward passes and optimizers.
• Core.TypedGraph: reusable shape-indexed graph wrappers.
• Core.Functional: operation-generic Ops interface and curried syntax.
• Core.Trainer.Types: scalar trainer contracts without backend construction.
• Core.Trainer: construct scalar trainers; each training step records an eager tape.
• Core.Trainer.EagerOps / GraphOps: Ops instances that record an eager tape or a typed graph. Typed graphs feed the autograd transforms and proofs.

This facade re-exports the complete runtime. Library code should import the part it uses: Functional.Ops for generic operators, Trainer.Types for trainer consumers, and the relevant Session submodule for eager internals.