MokuMoku Town 🏙️
A virtual office in your browser. Loosely connected, deeply focused.

LightGBM-MoE 🌲
A regime-switching / Mixture-of-Experts extension of LightGBM.

• #7247 — Added LightGBM-MoE to the official external repositories list

  • Documented a C++-native Mixture-of-Experts extension that combines specialized GBDTs through a learned gating function.
  • Why it matters: It makes regime-aware gradient boosting easier to discover for problems where one global model struggles with heterogeneous data, such as changing market conditions or distinct user segments.

• #7246 — Added native int8 input support for pre-discretized features (under review)

  • Eliminates the intermediate float32 copy at the Python–C++ boundary, reducing feature-matrix memory usage by up to 75% while preserving identical predictions.
  • Why it matters: Large, low-cardinality datasets can be passed to LightGBM without temporarily quadrupling their memory footprint, making training and inference more practical in memory-constrained environments.

• #2222 — Removed unnecessary deep copies in module sampling and batch-shape promotion
  • Replaced full parameter and distribution copies with structure-only or shallow copies, substantially lowering peak memory usage and making eager batch-shape promotion about 3× faster on large arrays.
  • Why it matters: Bayesian inference with large neural networks and distributions can avoid duplicating gigabytes of unchanged parameters, reducing out-of-memory failures and improving eager-execution performance without changing model behavior.