moe: offloaded models are never removed from the expert registry #173

Problem

_LAYERS only grows and each OffloadedLayer holds all tier handles (host expert tensors and GPU slots). OffloadReport.release() frees leases but not layers, so reloading a model in a server keeps the old experts in host and GPU memory.

Evidence and reproduction

opt_prime/opt_prime/accel/moe/ops.py:20-29. Ran: apply -> release -> del -> gc three times gave 2 -> 4 -> 6 registered layers.

Environment

optimus-rm @253be20; Python 3.10, torch 2.10.0+cu128, transformers 5.1.0; RNGD+A100 host (CPU / one A100 for the checks).

Scope

moe/ops.py, OffloadReport.

Done when

release() unregisters its layers; a test checks the registry is empty after release.

Owner

@gspark-etri