[brndnmtthws]

What changed

Two targeted CPU optimizations, with the existing numerical contracts intact:

• Build dense chain-order histograms only for the tree's sampled feature pool. Sweep up to four selected columns together, avoid per-feature scheduling on small nodes, and subtract only the pool's sibling bins. Level/node sampling stays inside that pool. Sparse and fixed-block builds retain the existing path; other backends inherit a full-build fallback.
• Route margin batches below 16 rows through the existing optimized RowForest used by single-row APIs, rather than the older CompactForest row traversal. Preserve per-row base margins, iteration ranges, DART weights and output order. Wide CSR and shrinkage paths stay unchanged.

No SIMD changes, floating-point reassociation, public configuration changes or model-format changes. Default full-feature training is not the target of the histogram optimization.

Before / after

Baseline: f6c5d8f (v0.4.1).
Rust 1.99.1 / LLVM 23.1.1, release profile, default features, no extra RUSTFLAGS.
Shared virtualized x86-64 cloud host exposing AMD EPYC 9V74 and nine logical CPUs. This is not a dedicated 80-core host, and these results do not generalize to other hardware.

Medians from serial baseline/candidate/candidate/baseline runs, ten samples per binary. Each process discards a warmup. Small-batch samples average 1,000 calls. Full raw samples, min/max ranges, binary hashes and reproduction instructions are in the benchmark report.

Workload	Threads	Before	After	Speedup
Warm hist training, 20k×128, 30 rounds, colsample=0.25	1	262.26 ms	106.79 ms	2.46×
Same	4	144.95 ms	71.24 ms	2.03×
Same, fresh matrix + binning included	1	347.47 ms	204.65 ms	1.70×
Same	4	164.30 ms	101.25 ms	1.62×
Predict 1 row, 100 trees	1	1.222 µs	0.699 µs	1.75×
Predict 8 rows, 100 trees	1	9.995 µs	5.308 µs	1.88×
Predict 15 rows, 100 trees	1	18.492 µs	9.919 µs	1.86×

Synthetic deterministic regression data, depth 6, 256 bins, eta=0.1. Small-batch models train on 10k×30 rows. Warm training reuses cached bins; cold includes DMatrix construction and binning. Prediction excludes model training and lazy layout construction. Serialized-model and prediction checksums agree across both binaries, every repeat and thread count.

Controls and limitations

Full-feature training and 16-row batches are near parity. Initial four-thread large-batch prediction and 80%-feature training runs were noisy and showed overlapping slowdowns. Both initial results are retained. A 50-sample-per-binary follow-up gave:

Control, 4 threads	Before	After
Large-batch prediction, 100k rows	10.78 ms	10.25 ms
Full-feature training	118.96 ms	117.44 ms
80%-feature wide training	172.93 ms	143.46 ms

No large-batch prediction win is claimed. Broad-pool training is less stable than the 25%-feature case. Missing-data fallback is also measured in the full report. Single-row APIs already used RowForest, so this does not claim to speed those up. An eight-lane x86 batch-walk experiment regressed and was dropped.

Validation

• 963 tests/doctests passed with RAYON_NUM_THREADS=4 cargo test --profile ci --all-features; five intentionally ignored tests remain ignored.
• All-target/all-feature clippy with warnings denied, rustdoc with warnings denied, root/fuzz/CUDA rustfmt, and the runner's Ruff/ty checks passed.
• New bitwise histogram tests cover NaNs, U8/U16 columns, irregular/repeated rows, blocked/sparse fallback, default backends, poisoned unused bins and thread/growth-policy equivalence.
• Small-batch tests cover dense/borrowed/sentinel/CSR/gathered input, base margins, iteration ranges and existing scalar/multiclass/vector/linear/DART model cases.
• Independent correctness review found no blockers.

GPU-dependent tests skip without a usable device. No actual GPU execution, Arm runtime measurement, Python-extension rebuild or external XGBoost/LightGBM fixture generation was done locally. CI passed all 25 jobs on published head a74a7ec35e1fea7c2b82971ad1378bfb7432047c, including Rust tests on x86 Linux/Arm Linux/macOS, MSRV, semver compatibility, XGBoost/LightGBM parity, Python wheels and sdist, CUDA build checks, and fuzz smoke.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-22T14:28:12.328982Z	a74a7ec	Draft marked ready
🔒 Security Review	✅ Completed 2026-10-22T14:28:59.438286Z	a74a7ec	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.