idun

Tabular learning on the Apple silicon GPU. idun trains three learners on Metal, each with a CPU reference backend:

• idun_gbm (idun.gbm): histogram gradient boosting with LightGBM's model (depth-wise or leaf-wise trees, quantized gradients, categorical splits, seed bags). Its CPU and Metal backends build the same model, bit for bit.
• idun_tabm (idun.tabm): TabM (Gorishniy et al., ICLR 2025), a parameter-efficient ensemble of MLPs, for classification, regression, and HL-Gauss regression (a target's Gaussian-smoothed bins).
• idun_realmlp (idun.realmlp): RealMLP-TD (Holzmüller et al., NeurIPS 2024).

Every fit is deterministic: the same data, parameters and seed give the same bits, run after run (the networks on the same GPU). The estimators follow scikit-learn's conventions without depending on it.

idun is in alpha (0.2): the interfaces may still change.

Install

pip install idun            # Apple silicon with macOS 15 or later: the Metal wheel; elsewhere the CPU-only wheel
pip install idun-numerai    # the Numerai port

Python 3.12–3.14; the only dependencies are NumPy and numba. Metal needs an Apple silicon Mac (M1 or newer) with macOS 15 or later. On Linux and Intel Macs the same package installs without the Metal library and runs the CPU backends; asking for Metal there raises idun.errors.BackendUnavailableError rather than falling back.

Quick start

import idun
from idun.gbm import GBMRegressor
from idun.realmlp import RealMLPRegressor
from idun.tabm import TabMClassifier

gbm = GBMRegressor(num_trees=300, learning_rate=0.05).fit(X_train, y_train)
pred = gbm.predict(X_test)
cpu = GBMRegressor(num_trees=300, learning_rate=0.05, backend="cpu").fit(X_train, y_train)  # the same trees

clf = TabMClassifier(categorical=["sector"]).fit(X_train, labels)  # any labels, two classes or more
proba = clf.predict_proba(X_test)

mlp = RealMLPRegressor(n_seeds=4).fit(X_train, y_train)  # a bag of four seeds; predicts their mean
each = mlp.predict(X_test, per_seed=True)  # (n, 4): each seed's prediction

gbm.save("gbm.idun")
same = idun.load("gbm.idun")

X is a 2-D NumPy array, a pandas DataFrame or an Arrow table; missing values are allowed. The backend is "metal" by default; pass backend="cpu" (idun_gbm) or backend="reference" (the networks' float64 reference) to run on the CPU. Nothing switches backends behind your back.

Below the estimators: idun.Dataset bins a table once for reuse; the functional layer (idun.gbm.train, idun.tabm.train_tabm, idun.realmlp.train_realmlp, seed bags through train_seeds and predict_mean) trains on binned data or on idun.data.PackedDataset, a memory-mapped, date-sorted panel for data larger than memory, which Metal reads without a copy. The docstrings document every public name.

How it compares

On five OpenML tasks on an Apple M1 (8 GB): idun_gbm's RMSE is within 3 % of the median of LightGBM, XGBoost and scikit-learn's HistGradientBoosting at the same nominal settings; idun_tabm and idun_realmlp are within 2 % of their reference implementations (the tabm package and pytabkit on Torch's MPS device) on every task and better on most, and idun_tabm fits in 0.45 of the reference's time.

Licence

Apache-2.0. The NOTICE file lists the work idun builds on, including MLX's Steel GEMM (MIT), whose structure the networks' GEMM kernel adapts.