# infer/weight-stream — PLAN12 Track 14: layer-wise weight streaming (AirLLM model).
# Runs a model whose full weights exceed memory by keeping one decoder layer resident
# at a time (load→compute→evict). Each concern owns its own subdir; new ones drop in here.
add_subdirectory(arch-compat)   # 14.04 — per-arch layer-layout descriptor + compat gate
add_subdirectory(checkpoint-split) # 14.02 — checkpoint -> per-layer shards (header parse + grouping)
add_subdirectory(persist) # 14.03 — ModelPersister abstraction, LRU cache, prefetch worker
add_subdirectory(engine) # 14.01 — layer-wise streaming forward (load→compute→evict)
add_subdirectory(compress) # 14.05 — optional offline block-quant compression (Int8/NF4)
