# infer — MLX C++ forward-pass engine: sampler, penalties, KV/cache substrate,
# grammar/JSON, generate. Owns its direct files; model-load/ and spec/ own theirs.
mlxm_sources()
add_subdirectory(model-load)
add_subdirectory(moe-async)    # PLAN12 12.04 — async expert I/O pipeline + router-lookahead prefetch
add_subdirectory(moe-pin)      # PLAN12 12.02 — hot-store pinning, usage history, heat-decay repin
add_subdirectory(moe-stream)   # PLAN12 12.01 — expert streaming policy (LRU, batch-union, config)
add_subdirectory(gdn-segment)  # PLAN12 15.03 — segmented GDN prefill driver (state threading + config)
add_subdirectory(mla)          # PLAN12 11.10 — compressed MLA-style KV cache (latent + rotary, absorb)
add_subdirectory(perf-toggle)  # PLAN12 12.11 — byte-identical perf-toggle discipline (neutrality proof)
add_subdirectory(ram-budget)   # PLAN12 12.06 — RAM-budget autotune: size the cache to the machine
add_subdirectory(weight-stream) # PLAN12 Track 14 — layer-wise weight streaming (load→compute→evict)
add_subdirectory(spec)
