TorchLean API

NN.Backend.Planner

Backend Planner #

The planner is the bridge between a semantic graph and backend capsules.

At this stage it is deliberately small: given an execution config, an operation tag, and a capsule registry, choose an admissible capsule or explain why none is available. Graph-aware layers can recover those operation tags from NN.IR.OpKind, lower adjacent nodes, and eventually produce executable command buffers without changing the contract story.

A backend choice for one graph operation or fused operation.

Instances For

    A simple execution plan: one selected capsule per requested operation.

    Instances For

      Names of the selected backend capsules, useful for audits and logs.

      Instances For

        Whether every selected capsule is admitted by the config's trust/device/backend policy.

        Instances For

          Put preferred-provider capsules first while preserving the relative order otherwise.

          Instances For

            Choose a backend capsule for one operation.

            Instances For

              Choose backend capsules for a sequence of operations.

              Instances For

                Choose backend capsules after filtering the registry by machine/build availability.

                Instances For