Skip to content

CTBoost 0.1.56

CTBoost 0.1.56 is an alpha release focused on training throughput, CUDA and distributed foundations, and portable inference. It does not change the conditional-inference learner: the protected split-statistics sources are byte-identical to 0.1.55. The quadratic feature test with no multiplicity adjustment remains the default, grouped testing remains opt-in, and Bonferroni remains a stopping-threshold correction rather than a score optimizer.

Highlights

  • Sparse CSC quantization now scans each column linearly, and CPU node histograms can run deterministically across features.
  • Pool.from_cuda_quantized(...) accepts validated, single-device CUDA Array Interface buffers without a host-data round trip.
  • GPU training uses the shared conditional-inference candidate selector, including grouped statistics and per-node Bonferroni adjustment, while split optimization continues to use the original raw bins.
  • The Spark adapter adds fail-closed barrier execution, stable global query IDs, and explicit collection rather than silently collecting in mode="auto".
  • Version-2 JSON predictors can carry a fitted Python preprocessing pipeline; loaders now reject oversized, duplicate-key, non-finite, cyclic, out-of-bounds, and structurally inconsistent artifacts.
  • Prepared-numeric, inference-only source runtimes are included for Java 17 and R, backed by a shared conformance corpus.

Compatibility and boundaries

The Python runtime reads valid version-1 and version-2 JSON predictors. Older version-1-only readers reject new version-2 artifacts. Raw fitted preprocessing runs only in Python; the R and JVM modules require prepared numeric features and are source modules in this release, not CRAN or Maven publications.

Spark training now defaults to fail-closed mode="auto"; pass mode="collect" explicitly to permit driver collection. The initial distributed GPU transport remains TCP/barrier based and is not an NCCL, GPU-direct, elastic, or fault-tolerant multi-node stack.

The native Pool C++ class layout changed. Ordinary wheel users receive a matching rebuilt extension, but applications that compile or link CTBoost's repository-native headers or static core must rebuild. Generated model-specific C/C++ scorer interfaces are unchanged.

Model fingerprints detect accidental inconsistency; they are not cryptographic signatures. Exported models remain trusted artifacts. Valid legacy model and pipeline states remain supported, while malformed states that older versions tolerated may now fail closed.

Benchmark interpretation

The throughput work is intended to preserve deterministic model behavior, not to claim a predictive-score improvement. The latest measured TabArena evidence remains the 0.1.55 three-dataset default smoke at 1054.0 provisional Elo. It is not a full or official leaderboard result and is not evidence of a 0.1.56 Elo gain.