sim-lib-compute-auto 0.3.0

Automatic tensor compute site selection for SIM.
Documentation

sim-lib-compute-auto

Automatic compute-site selection for SIM tensors.

AutoTensorExecutor accepts bounded profile records captured through the current synthetic harness or later physical-device producers. Profiles record adapter, driver, backend, limits, transfer and allocation evidence, tile sizes, operation sample distributions, thermal and power context, evidence kind, observed identity, and provenance. They persist through a caller-supplied Table or Dir value with bounded keys and record sizes; the auto library never opens a host path.

Automatic placement stays conservative: absent, stale, incompatible, or inconclusive evidence routes to the published CPU tensor executor from sim-numbers. Fresh compatible evidence can select the modeled resident provider only when verify_physical accepts a physical-device profile whose claimed identity still matches the observed identity. Current synthetic and host-emulated records route to CPU, and each execute/flush records provider choice, materialization bytes, and synchronization counts in the routing ledger.