bonsai-gbt 1.5.0

Histogram gradient-boosted trees with a C++23 core

A histogram gradient-boosted tree library and CLI in modern C++23.

Documentation  ·  Install  ·  Guide  ·  Decisions  ·  Releases

---

What is bonsai?

bonsai is a from-scratch, histogram-based gradient boosted trees (GBT) library and command-line tool written in C++23. It pairs a small, concept-checked component API (objectives, growers, split finders, samplers) with compile-time dispatch in the training hot path, and ships the benchmark harness that pits it against XGBoost, LightGBM, and CatBoost on real data. The aim is a readable, thoroughly documented GBT: a reference-grade implementation that competes with the production libraries instead of merely tolerating comparison with them.

• Compile-time dispatch, concept-checked components. The runtime TOML config resolves once to a monomorphized Booster<Objective, Grower, Splitter, Sampler>; no virtual calls in the hot path, and contract violations fail at compile time. Adding a component is a short recipe.
• Five growers, one engine. depthwise (XGBoost-style), leafwise (LightGBM-style), oblivious (CatBoost-style), and their CUDA twins cuda_depthwise / cuda_oblivious; with 7 objectives and 3 samplers the dispatch space is 105 statically-typed combinations, selectable per run from config.
• Deterministic parallelism. Models are bit-identical across runs, thread counts, and even CPU architectures (arm64 == x86-64), a property no reference library offers, enforced per-commit in CI (decisions 49/59/60).
• A guide, not just docs. The guide explains gradient boosting chapter by chapter: concept, math, then the ~50 real lines that implement it here, then an experiment against the reference libraries.

Install

pip install bonsai-gbt --find-links https://github.com/daniel-m-campos/bonsai/releases/expanded_assets/v1.4.0

Wheels cover Linux x86_64/aarch64 and macOS arm64, Python 3.9 to 3.13, no toolchain needed. The linux x86_64 wheel is CUDA-enabled at 2.3MB total: GPU training works out of the box on any NVIDIA driver R525+, it behaves exactly like a CPU wheel on machines without a GPU, and every release's CUDA wheel passes a live GPU validation before it ships (decision 70). The full story, docker image included, is Install; everything past a wheel (the CLI binary, development setups, CUDA source builds) is Building from source.

Quick start

import bonsai

model = bonsai.BonsaiRegressor(
    n_iters=200, learning_rate=0.05, grower="leafwise",
    early_stopping_rounds=20,
    params={"tree.lambda_l1": 0.5},   # any dotted config key the CLI accepts
)
model.fit(X_train, y_train, eval_set=(X_valid, y_valid))
pred = model.predict(X_test)
model.save("model.msgpack")           # loadable by `bonsai predict` and vice versa

The CLI (a source-build artifact) drives the same engine with the same keys and the same models:

bonsai fit      -c CONFIG --model OUT.msgpack
bonsai predict  -c CONFIG --model IN.msgpack [--data CSV] --out PREDS.csv
bonsai eval     -c CONFIG --model IN.msgpack [--data CSV]
bonsai info                        # list (objective, grower, sampler) combos
bonsai params                      # dump the default config as TOML

Any key overrides inline (bonsai fit -c config.toml --set tree.max_depth=8 --set dispatch.grower_name=oblivious ...), and make fit-benchmark trains and times bonsai against xgboost/lightgbm/catboost on California Housing in one command. The rest of the API is one read: the API tour.

Performance

Two divisions, per the benchmark charter: quality (accuracy, timing never citable) and perf (latency and memory, quality as a sanity guard). The complete evidence, generated from every committed results file, is the results ledger.

Quality: the external standings

On the Grinsztajn et al. tabular benchmark, 55 OpenML tasks selected by third parties, three seeds, matched campaign knobs, best variant per library (decision 68, evidence):

library	mean rank	outright wins
bonsai	1.44	36
lightgbm	2.51	5
xgboost	2.84	6
catboost	3.22	8

The one knob that translates ambiguously between libraries is bracketed: under XGBoost's own min_child_weight=1 convention the top two swap on mean rank (2.11 vs 2.04) while bonsai keeps the most second-or-better finishes; both runs are in the ledger. Reproduce either: pip install bonsai-gbt[bench], then python -m bonsai.bench.grinsztajn out.jsonl --report.

Perf: fit seconds at scale

From the same-pod 2026-07 re-baseline (dual-EPYC-9554 host with an L40S; fit() timed end to end including each library's own ingest; test r² in parentheses; raw runs):

rows	bonsai cuda dw	bonsai cuda obl	xgb cuda	catboost gpu	lgbm cpu	bonsai cpu obl
250k	0.5s (.871)	1.0s (.875)	0.8s (.872)	1.6s (.875)	2.5s (.872)	5.2s (.875)
1M	1.1s (.876)	1.4s (.876)	1.7s (.876)	2.3s (.876)	5.0s (.877)	7.3s (.876)
4M	4.5s (.878)	4.4s (.875)	5.3s (.878)	5.0s (.877)	19.9s (.879)	20.2s (.875)
16M	20.5s (.879)	18.4s (.876)	19.9s (.880)	18.5s (.876)	111.3s (.879)	73.3s (.876)

Honest caveats, because benchmarks without them are advertising: identical-model GPUs across the rental fleet measure up to ~25% apart, so only same-pod columns compare. bonsai owns the fastest slot at every row scale, edging CatBoost at 16M (18.4 vs 18.5s, both .876) and beating XGBoost-GPU (19.9s); on wide data CatBoost keeps the lead, with bonsai second (the cols-scaling table is in the ledger). Off the fixed-iteration axis, the 16M accuracy-vs-time frontier now belongs to bonsai: fastest to every measured accuracy up to ~.895 r², a statistical tie with CatBoost through the .897-.898 plateau, and the measured ceiling (.8981); CatBoost's remaining edge is a cheaper marginal round that pays off only past ~450 rounds at this scale (decision 72, chart in the ledger). Peak host RSS at 16M is 7.0GB vs XGBoost's 22.2GB and CatBoost's 19.4GB, and predict is ~3x faster. Two earlier apparent gaps against CatBoost were bonsai bugs, since fixed (decisions 63/64); the path from 3x behind to this table is guide chapter 11.

Claims and proofs

Every claim links a reproducible run and the decision that records it; the point of a small, measured library is that you can check it.

Claim	Evidence
Bit-identical models across CPU architectures (arm64 == x86-64) at a fixed thread count; no reference library offers this	decisions 59/60; asserted per-commit by cross-arch.yml via scripts/model_hash.py
Best mean rank on the 55-task Grinsztajn benchmark under either min_child_weight convention (36 outright wins; second-or-better on 50/55, never last)	grinsztajn-2026-07, decision 68
Fastest GPU slot at every row scale; at 16M oblivious edges catboost and beats xgboost-GPU at matched accuracy	rebaseline jsonl, scale-edge, decisions 62-64
The only GBT whose GPU path ships in a 2.3MB pip install, validated on live GPU hardware per release	decision 70; wheels.yml
Within ~8% of xgboost-hist at 16M rows on CPU, host-dependent: a dead tie on one pod, xgboost ahead on another	decision 61; prefetch-round jsonl
Best library on 9 of 10 datasets of the internal quality campaign	quality-campaign, decisions 56-57
Categorical parity with catboost within chance-band, via preprocessing not an engine feature	decision 58; categorical-tradeoff; encoding.py
~3x less host memory than xgboost at 16M (7.0 vs 22.2GB) and ~3x faster predict	rebaseline jsonl
Ranking is a measured, scoped gap: ~+0.015 NDCG@10 to a listwise loss, not pairwise LambdaRank	ranking-tradeoff; probe_ranking.py
Every feature earns its place by measurement; refutations are recorded too	the feature-admission gate; declines in decisions 58/62/67

Documentation

The home is daniel-m-campos.github.io/bonsai, four doors and a notebook:

• Learn: gradient boosting from intuition to the shipping code, one concept per chapter, each with an experiment against the reference libraries.
• Use: install, the API in one read, building from source.
• Lineage: what XGBoost, LightGBM, and CatBoost each contributed, and whether measurement adopted, rebuilt, or declined it.
• Method: the measurement discipline, portable beyond GBTs; its rules in the benchmark charter and its evidence in the results ledger.
• The engineering notebook: the decisions log and the architecture notes.

Repo-only records, unpublished by design: the project retrospective, the original proposal, and the context/roadmap.

Project layout

include/bonsai/   public headers (Booster, Tree, Grower, Sampler, …)
src/              implementation + CLI (src/cli/)
python/           the bonsai package (bindings, encoding, bonsai.bench)
tests/unit/       Catch2 unit + parity tests (ctest)
benchmarks/       evidence docs + committed results data
scripts/          uv-managed Python: compare.py, probes, render_results.py
configs/          example TOML configs
docs/             the documentation site source + design records

License

MIT © 2026 Daniel M Campos. See LICENSE.