Skip to content

Gradient Boosting

MPSBoost gradient boosting models use the in-project native tree engine. The CPU backend is the correctness oracle, and the MPS backend targets histogram hot paths that suit Apple GPU execution.

Regression

Primary entries:

  • GradientBoostingRegressor
  • MPSBoostRegressor

Supported capabilities:

  • squared error
  • quantile
  • Poisson
  • Tweedie
  • sample weight
  • monotonic constraints
  • interaction constraints
  • L1/L2 regularization
  • leaf-wise / level-wise growth
  • model save/load

Classification

Primary entries:

  • GradientBoostingClassifier
  • MPSBoostClassifier

Supported capabilities:

  • binary logistic
  • native CPU multiclass softmax
  • OvR compatibility strategy
  • predict_proba
  • decision_function
  • sklearn model selection

Multiclass defaults to native softmax when it is available on CPU. MPS requests use the observable compatibility strategy and report the actual strategy in the training summary.