# infer/mla — PLAN12 11.10: compressed MLA-style KV cache (latent + shared rotary key) and the
# two equivalent attention paths (non-absorbed reconstruct vs absorbed decode fast path).
# CPU-plain, arch-agnostic substrate; the MLX/Metal-backed pack that consumes it lands per-pack
# under src/arch/* once an MLA-family model is present. mlxm_sources() globs this dir's *.cpp —
# filesystem is the registry, so new MLA files compile with zero CMake edits.
mlxm_sources()
