Skip to content

MPSBoost 0.3.0 Release Audit

This document records the release gate for the 0.3.0 v2 arboretum milestone.

Scope

0.3.0 supports:

  • dense numeric regression;
  • binary and multiclass classification;
  • squared-error objective;
  • quantile, Poisson, and Tweedie regression objectives;
  • deterministic quantization;
  • depth-limited histogram trees;
  • decision trees, random forests, ExtraTrees, and CatBoost-like numeric estimators;
  • native CPU multiclass softmax with explicit OvR compatibility;
  • CPU-suitable isolation forest anomaly scoring;
  • pointwise learning-to-rank scoring with query-group validation;
  • real MPS gradient, histogram, split-scan, partition, and buffer-pool paths;
  • explicit CPU oracle mode;
  • model save/load;
  • feature importance, permutation importance, and controlled SHAP-like explanations;
  • import-time MPS environment guidance with copy-paste setup and skip commands;
  • cache diagnostics, explicit cache creation, and safe cache clearing.

Not included:

  • sparse matrices;
  • native MPS multiclass softmax;
  • official third-party SHAP TreeExplainer integration;
  • categorical model persistence;
  • public GPU prediction;
  • full third-party API compatibility.

License

  • Project license: Apache-2.0.
  • Runtime dependency: NumPy, with permissive license expression reported by package metadata.
  • Build/test-only dependencies are not bundled into runtime wheels.
  • The wheel must include the project LICENSE file.

Wheel Content Rules

The release wheel must contain only runtime package files:

  • Python package files;
  • the native extension;
  • the compiled Metal shader library;
  • package metadata and license metadata.

The release wheel must not contain:

  • specs/;
  • tests/;
  • benchmarks/;
  • .github/;
  • build directories;
  • cache directories;
  • raw .metal, .air, or temporary shader files;
  • credentials or runner files.

The native extension may link to macOS system libraries and frameworks required for Python, C++, Objective-C runtime, Foundation, CoreFoundation, and Metal. It must not link to heavyweight ML runtimes or private project-local absolute paths.

Validation Matrix

Required before publishing 0.3.0:

  • local full test suite;
  • GitHub hosted CPU/package tests for Python 3.10 and 3.13;
  • self-hosted real Metal GPU tests for Python 3.10 and 3.13;
  • twine check for the exact uploaded wheels;
  • fresh PyPI install and real MPS smoke test.

Benchmark Evidence

Checked-in benchmark results:

  • benchmarks/results/s4-m2-ultra-py313.json
  • benchmarks/results/s4-m2-ultra-py313.md
  • benchmarks/results/s6-m2-ultra-py313.json
  • benchmarks/results/s6-m2-ultra-py313.md

The S6 report records both GPU wins and small-data regressions. The gbdt-large-wide end-to-end scenario reached a 1.629x median speedup on the recorded M2 Ultra validation machine.