skmetal 0.3.3

GPU-accelerated scikit-learn via Apple Metal

skmetal

Apple Silicon GPU acceleration for scikit-learn

import skmetal
from sklearn.linear_model import LinearRegression

@skmetal.accelerate
def model():
    return LinearRegression()

m = model()
m.fit(X_train, y_train)
m.predict(X_test)

---

Overview

skmetal executes scikit-learn estimators on Apple Silicon GPUs via Metal Performance Shaders and custom Metal compute kernels. Decorate any function that returns an estimator with @skmetal.accelerate and fit()/predict() run on the GPU — no code changes required.

Apple Silicon's unified memory architecture enables zero-copy data sharing: numpy arrays are passed directly to Metal via bytesNoCopy, eliminating data transfer overhead.

---

Requirements

• macOS 14+
• Apple Silicon (M1-M5)
• Python 3.10-3.12
• Swift 6.1 (xcode-select --install)
• scikit-learn >= 1.5

---

Installation

pip install skmetal

Note: macOS 14+ and Apple Silicon required. Xcode not needed for the pip package.

From source

git clone https://github.com/abderahmane-ai/skmetal.git
cd skmetal
bash build.sh
pip install -e .

---

Usage

Decorator (recommended)

import skmetal
from sklearn.linear_model import LinearRegression

@skmetal.accelerate
def model():
    return LinearRegression()

m = model()
m.fit(X, y)
m.predict(X_test)

The decorator also works with pipelines:

@skmetal.accelerate
def pipeline():
    return Pipeline([
        ("scaler", StandardScaler()),
        ("clf", LogisticRegression()),
    ])

pipe = pipeline()
pipe.fit(X, y)
pipe.predict(X_test)

Function call

model = skmetal.accelerate(LinearRegression())
model.fit(X, y)

Context manager

with skmetal.accelerate_context():
    model = LinearRegression()
    model.fit(X, y)

---

Supported Estimators

Estimator	GPU Strategy	Speedup
LinearRegression	Normal equations via MPS GEMM	5.93x
Ridge	Fused centering + XTX + XTy (1 dispatch)	1.16x
LogisticRegression	IRLS (3-5 Newton iterations, fused)	0.91x
Lasso	Coordinate descent + GPU residual updates	--
ElasticNet	Coordinate descent + GPU residual updates	--
TruncatedSVD	Randomized SVD, no centering (all BLAS-3)	2.53x
KMeans	Single fused command buffer (all iterations on GPU)	0.69x
DBSCAN	GPU pairwise distance + per-point neighbor counting	--
GaussianNB	GPU mean/var per class	--
StandardScaler	Fused Welford (1 dispatch)	8.27x
MinMaxScaler	Column min/max with threadgroup tree reduction	--
RobustScaler	GPU quantile approximation	--
KNeighborsClassifier	GPU pairwise distance + fused voting	--
KNeighborsRegressor	GPU pairwise distance + fused averaging	--
NearestNeighbors	GPU pairwise distance + index	--
HistGradientBoostingRegressor	C++ HGBT from sklearn (no custom GPU)	--
HistGradientBoostingClassifier	C++ HGBT from sklearn (no custom GPU)	--

Estimators below 1.0x speedup are dispatch-limited at n <= 50K. Speedup improves to 2-5x at n >= 500K where compute dominates overhead.

---

Architecture

Zero-copy pipeline

numpy array -> np.ctypes.data -> UnsafeMutableRawPointer -> MTLBuffer(bytesNoCopy:) -> GPU
                   |                                                    |
                   +--------- same physical memory ---------------------+

Metal kernels (14 files)

Kernel file	Operations
ReductionKernels.metal	reduce_sum, reduce_mean_var (Welford)
ArgminKernels.metal	argmin_rows
KMeansKernels.metal	assign, partial_sum, combine, normalize, batch fused
KNNKernels.metal	tile top-k, merge, fused vote classify/regress
PairwiseDistKernels.metal	pairwise_distance_squared, pairwise_distance_direct
DistanceKernels.metal	row_norm_sq, compute_mindists, distance_correct
IrlsKernels.metal	irls_weight, scale_rows, compute_linear_irls, compute_error_scale, l2_reg_irls, multinomial_hessians
CenterColumns.metal	column_means, center_columns
ElementWiseKernels.metal	sigmoid, subtract, add_scalar, axpy, norm_sq, transpose_f32, row_max, row_sum, softmax, negate
ExtraKernels.metal	soft_threshold, column_transform, scale_f32, sv_init, sv_hook, sv_shortcut
StandardScalerKernels.metal	scaler_fit (fused Welford)
GemmKernels.metal	gemm_simple (fallback)
MinMaxKernels.metal	column_minmax (threadgroup tree reduction)
TreeKernels.metal	tree_predict, tree_predict_all

Swift bridge (47 C-callable functions)

All skmetal_* functions use @_cdecl for direct ctypes export. Every function accepts raw pointers.

Project structure

skmetal/
  skmetal/
    __init__.py
    _bridge.py           ctypes -> Swift (47 functions)
    _config.py           Config dataclass
    _dispatch.py         estimator registry + wrapping
    accelerate.py        @accelerate decorator + accelerate_context
    estimators/
      _base.py           BaseGPUEstimator abstract class
      _registry.py       Estimator registry (17 estimators)
      linear_model.py    LinearRegression, Ridge, LogisticRegression, Lasso, ElasticNet
      cluster.py         KMeans, DBSCAN
      decomposition.py   TruncatedSVD
      ensemble.py        HistGradientBoostingRegressor, HistGradientBoostingClassifier
      naive_bayes.py     GaussianNB
      neighbors.py       KNeighborsClassifier, KNeighborsRegressor, NearestNeighbors
      preprocessing.py   StandardScaler, MinMaxScaler, RobustScaler
    utils.py
  skmetal_bridge/        Swift + Metal
    Sources/SkMetalBridge/
      Bridge.swift       47 @_cdecl exports
      MetalContext.swift
      Kernels/*.metal    14 Metal kernel files
  benchmarks/
    run_compare.py       benchmark runner
    benchmark_suite.py   full suite
    baseline.json
  tests/
    test_correctness.py  18 estimator correctness tests
    test_dispatch.py     7 dispatch logic tests
    test_kernels.py      55 kernel unit tests
    test_config.py       13 config API tests
    test_accelerate.py   20 accelerate edge-case tests
    test_stress.py       27 edge-case/stress tests
    conftest.py          shared fixtures
  build.sh
  pyproject.toml
  .github/workflows/
    ci.yml               build + test + benchmark regression
    release.yml          PyPI publish on tag
  LICENSE                MIT

---

Tests

test_correctness   18/18 pass — estimator parametrizations + pipeline + device info
test_dispatch       7/7  pass — registry, wrapping, pipeline, decorator
test_kernels       55/55 pass — all 14 kernel files covered
test_config        13/13 pass — config API, dtype rejection, threshold, reset
test_accelerate    20/20 pass — decorator, context manager, edge cases
test_stress        27/27 pass — contiguity, shape mismatch, extremes, CPU fallback

---

Benchmarks

python benchmarks/run_compare.py

---

Development

git clone https://github.com/abderahmane-ai/skmetal.git
cd skmetal
bash build.sh
pip install -e ".[dev]"
pytest tests/
python benchmarks/run_compare.py

Adding a new estimator

1. Create skmetal/estimators/my_model.py with MetalMyModel(BaseGPUEstimator)
2. Register in _registry.py (GPU_ESTIMATORS + PIPELINE_PATTERNS)
3. Add module path in _dispatch.py (module_map)
4. Write a Metal kernel if needed
5. Add a parametrized test case in test_correctness.py

---

License

MIT License. See LICENSE.