TorchLean API

NN.Backend.Grouping

Backend Kernel Groups #

GraphKernelPlan chooses a capsule for each runtime-relevant IR node. A GroupedKernelPlan places adjacent choices with the same operation and complete capsule in one audit group, so a graph with twenty consecutive ReLU nodes produces one audit row instead of twenty. Grouping does not translate graph semantics, fuse operations, or claim that a group executes as one kernel launch.

One audit group, with source IR node ids retained for diagnostics.

Instances For
    @[instance_reducible]

    A node-level graph plan grouped for trust-boundary audit.

    Instances For

      Project the operation and capsule selected for this group.

      Instances For

        Whether a planned node can share this audit group without losing contract evidence.

        Matching dispatch identities alone is insufficient: a caller can supply capsules with the same name, operation, provider, and device but different trust or numerical policies.

        Instances For

          Append a node to an existing group, preserving source-node provenance.

          Instances For

            Source IR node ids covered by the grouped plan, in graph order.

            Instances For

              Selected capsule names, in group order.

              Instances For

                Project groups to the capsule rows consumed by the contract check.

                Instances For

                  Audit the selected backend boundaries of the grouped plan.

                  Instances For

                    Place one graph-planned node in its own group.

                    Instances For

                      Add one planned node, sharing the preceding group when the boundary is identical.

                      Instances For

                        Adjacent nodes with the same backend boundary share one audit group.

                        Instances For