overview

Section O of the Feature Backlog: a graph kept in a store and loaded where a query reaches; planned as its own package. Open rows only; a row that is done moves to Backlog Done 2026-11 with its id. Status, priority, ids and the column needs are explained in Feature Backlog, which also holds the dependency graph.

Sources: original to this repository; the sources of each row are in its source column.

A graph too big for memory, kept in a store and loaded where a query reaches. It is planned as its own package, lib/MyceliumStore.jl (working name), on top of Mycelium. Daniel, 2026-11-05: not only for language models but also for heterogeneous multi-omics or streaming time series (markets, sensors). Most of the computation can be avoided in most cases, while particular data is still needed many times.

idfeaturestatusneedsprioritysourcewhy
O1Lazy factor nodes: a stub with a store key that loads its factor and parameters on first use (a message, a free energy), with an LRU cache and eviction; the graph’s structure stays residentplannedO2, O3mediumDaniel, 2026-11-05a model whose parameters do not fit in memory; mixture-of-experts serving does this for its experts (offloading, memory-mapped weights)
O2Nodes stored as description plus arrays: a registered type tag and version, the constructor’s arguments as data, and ps/st in a memory-mappable format (Arrow, SafeTensors, HDF5/JLD2 datasets, Mmap); no serialised functions or closures (Julia’s Serialization is not a stable format, and deserialising untrusted data is unsafe)planned—mediumDaniel, 2026-11-05factors are already split the Lux way, a small description and arrays; only init_* closures hold code, and they are needed only for initial parameters
O3A store interface with back ends: files keyed by node id, LMDB (memory-mapped, read-heavy), SQLite, RocksDB, Redis, object storesplannedO2lowDaniel, 2026-11-05one interface; the choice per deployment
O4The subgraph a query reaches: from the clamps and the query variables, the factors connected through unobserved variables (Bayes-ball / d-separation), and a schedule restricted to it; incremental appends (iSAM-like) touch only the newest factorsplanned—mediumDaniel, 2026-11-05; LenticulumLM’s Incremental Next-Token Predictionwhat makes lazy loading pay; full loopy inference would load everything
O5Parameters by key on demand instead of one graph-wide NamedTuple (the integer-slot store of the inference section), and training on a loaded subgraph with write-back of parameters and optimiser stateplannedI5, O1lowDaniel, 2026-11-05; ChatGPT reviewgradients are local per parameter key, so a loaded subgraph trains as a block of coordinates