# infer/weight-stream/compress — PLAN12 14.05: optional offline block-quant
#   compression for per-layer shards. Int8 (blocksize 2048, ~2x) and NF4
#   (blocksize 64, ~4x). CPU-only codec, no GPU dependencies. Mutually
#   exclusive with 14.03 prefetch (documented, not silent).
# mlxm_sources() globs this dir's .cpp into mlx-magic-lib (filesystem IS the registry).
mlxm_sources()
add_subdirectory(producer) # 14.05H — the WRITE side: 14.02 split -> .qz shards
