"""LPU V7 Accelerator Controller — the standing decision brain. This subpackage is NEW in V7 (mission §2.B/§2.E). It runs torch-free inside the accelerator server process, continuously reads runtime facts (process discovery, GPU samples, session telemetry) and: * classifies bottlenecks with measured evidence (bottleneck.py) * governs parameter changes across the 3 safety levels and the hot-reload / safe-boundary / restart-required classes (parameters.py) * runs the propose -> validate -> apply -> ack -> rollback lifecycle (optimizer.py) * orchestrates everything in a background service loop (service.py) Design rules inherited from the V6 core: * no fabricated metrics - anything unmeasured is labeled an estimate * read-only towards foreign processes; control ONLY via the session command channel (allowlisted, envelope-versioned, acked) * stopping the accelerator NEVER stops training """ from .parameters import PARAM_SPECS, ParamLevel, ChangeClass, ParamSpec from .bottleneck import BottleneckEngine, BottleneckReport from .optimizer import OptimizerEngine, Proposal, ChangeRecord from .service import AcceleratorService __all__ = [ "PARAM_SPECS", "ParamLevel", "ChangeClass", "ParamSpec", "BottleneckEngine", "BottleneckReport", "OptimizerEngine", "Proposal", "ChangeRecord", "AcceleratorService", ]