statgpu

GPU-accelerated statistical methods with sklearn-compatible API.

Documentation

• English docs: docs/en/ — full documentation index
• Chinese docs: docs/ — 中文文档
• Quickstart: Quickstart
• GLM + Penalty: Generalized Linear Model — 7 families × 10 penalties × 3 backends
• Cross-Validation: Cross-Validation Guide — PenalizedGLM_CV, LassoCV, RidgeCV
• Loss × Penalty × Solver Framework: Framework Guide — complete architecture, dispatch logic, coverage matrix
• Solver-Penalty Matrix: Solver × Penalty — solver dispatch and penalty routing
• Device & Memory: Device and GPU Memory
• PyTorch Backend: PyTorch Backend
• Distribution API: Distribution API — 15 distributions across 3 backends
• Multiple Testing: Multiple Testing — p-value adjustment and combination
• Changelog: Changelog

Features

• 🚀 3 Backends: NumPy (CPU), CuPy (CUDA), PyTorch (CUDA) — automatic device selection
• 🔧 sklearn-compatible: fit/predict/score API, sklearn.base.clone() supported
• 📊 GLM + Robust + Quantile + Cox: 10+ loss types (quantile, huber, bisquare, fair, cox_ph + 7 GLM families)
• 🔥 10 Penalties: l1, l2, elasticnet, scad, mcp, adaptive_l1, group_lasso, group_mcp, group_scad
• ⚡ 8 Solvers: exact, newton, lbfgs, irls, fista, fista_bb, proximal_irls_cd, proximal_newton — solver="auto"
• 🧮 Inference: penalized sandwich (L2) + oracle (SCAD/MCP) for Hessian-equipped losses; analytical Hessian for ordered models; kernel + bootstrap for quantile regression; debiased Lasso + simultaneous CI — GPU-native across NumPy/CuPy/Torch
• 📈 Nonparametric: KDE, kernel regression, B-splines, GAM
• 🧬 Unsupervised: PCA, KMeans, DBSCAN, GMM, UMAP, t-SNE, NNDescent (12+ classes)
• 📐 Distributions: 15 distributions across 3 backends via get_distribution() — API docs
• 🧪 Multiple Testing: adjust_pvalues + combine_pvalues + permutation_test
• 🔥 Cross-Validation: PenalizedGLM_CV (all losses × 10 penalties), RidgeCV, LassoCV, ElasticNetCV

Implemented Methods

Full method list with solvers, penalties, link functions →

Category	Classes	Highlights
Regression & GLM	13 classes	LinearRegression, Ridge, Lasso, ElasticNet, Logistic, Poisson, Gamma, InvGauss, NB, Tweedie, QuantileRegression, Ordered models (logit/probit, GPU inference)
Penalized GLM	11 classes	PenalizedGLM + 7 family wrappers + PenalizedQuantileRegression, PenalizedRobustRegression, PenalizedCoxPHModel × 10 penalties × 8 solvers
Cross-Validation	6 classes	RidgeCV, LassoCV, ElasticNetCV, LogisticCV, PenalizedGLM_CV, CoxPHCV
ANOVA	2 functions	f_oneway, f_twoway — GPU-accelerated
Covariance	3 classes	EmpiricalCovariance, LedoitWolf, OAS
Panel Data	2 classes	PanelOLS, RandomEffects
Nonparametric	5 classes	KernelRidge, KernelRidgeCV, pairwise_kernels, bspline_basis, natural_cubic_spline_basis
Semiparametric	1 class	GAM (penalized B-splines + GCV)
Unsupervised	12 classes	PCA, SVD, NMF, UMAP, t-SNE, KMeans, DBSCAN, GMM, AgglomerativeClustering
Survival	1 class	CoxPH (Breslow/Efron ties, robust SE)
Feature Selection	2 functions	fixed-X / model-X knockoff filters
Multiple Testing	3 functions	adjust_pvalues, combine_pvalues, permutation_test

Installation

# CPU only
pip install statgpu

# With GPU support (choose by CUDA major version)
# CUDA 11.x runtime:
pip install statgpu[gpu11]

# CUDA 12.x runtime:
pip install statgpu[gpu12]

# With PyTorch backend (CUDA 11.x)
pip install statgpu[torch]

# Development
pip install statgpu[dev]

# Formula interface
pip install statgpu[formula]

Quick Start

import numpy as np
from statgpu.inference import norm, poisson
from statgpu.linear_model import LinearRegression, PenalizedGLM_CV
from statgpu import adjust_pvalues, combine_pvalues

# Generate data using statgpu distributions (scipy-compatible API)
X = norm.rvs(size=(10000, 100))
y = X @ norm.rvs(size=100) + norm.rvs(size=10000) * 0.5

# Linear regression with GPU
model = LinearRegression(device='cuda')
model.fit(X, y)
print(f"R²: {model.score(X, y):.4f}")

# Penalized GLM with cross-validation
y_pois = poisson.rvs(mu=np.exp(X[:, :5] @ np.ones(5) * 0.1), size=X.shape[0])
cv_model = PenalizedGLM_CV(
    loss="poisson", penalty="elasticnet", l1_ratio=0.5,
    cv=5, device="cpu",
)
cv_model.fit(X[:, :5], y_pois)
print(f"Best alpha: {cv_model.alpha_:.4f}")

# Multiple-testing correction
reject, pvals_adj = adjust_pvalues(np.array([0.003, 0.02, 0.5]), method='bh')

# Global p-value combination
stat, p_global = combine_pvalues(np.array([0.01, 0.07, 0.03, 0.40]), method='fisher')

Device Control

import statgpu as sg

# Global setting
sg.set_device('cuda')  # Force GPU
sg.set_device('cpu')   # Force CPU
sg.set_device('auto')  # Auto-detect (default)

# Per-model setting
from statgpu.linear_model import LinearRegression
model = LinearRegression(device='cuda', n_jobs=4)

Benchmark Results (RTX 4090)

Full reports: results/unsupervised_bench_2026-06-27.md, results/glm_solver_benchmark_2026-06-23.md

Test environment: RTX 4090 (24GB), CuPy 14.1.0, PyTorch 2.8.0+cu128, scikit-learn 1.8.0, statsmodels 0.14.6, lifelines 0.30.3
Benchmark environment only; not installation requirements.

Real-Data Performance

Module	Dataset	n	p	Best Speedup	Precision
Poisson GLM	freMTPL2	678K	42	196.9x vs sklearn	coef_corr=1.000000
Gamma GLM	synthetic	678K	42	97.9x vs sklearn	coef_corr=0.9995
CoxPH	synthetic	1.9K	500	1.2x vs CPU	coef_corr=1.000
adjust_pvalues (BH)	synthetic	—	1M	0.55x	100% agreement
PenalizedPoisson(L1)	freMTPL2	678K	42	—	OK
PenalizedCoxPH(L2)	synthetic	1.9K	500	—	C-index match

Precision Summary

Module	Metric	Result
Poisson GLM	coef correlation vs sklearn	1.000000 (full freMTPL2)
Gamma GLM	coef correlation vs sklearn	0.9995
CoxPH	coef correlation vs lifelines	1.000
adjust_pvalues (BH)	reject agreement vs statsmodels	100% (100K to 5M p-values)
Penalized (L1/L2)	self-consistency	C-index match across penalties

Requirements

• Python >= 3.8
• NumPy >= 1.20
• CuPy (optional, for GPU; choose wheel matching CUDA major version)
  • CUDA 11.x: cupy-cuda11x
  • CUDA 12.x: cupy-cuda12x
• CUDA runtime compatible with selected CuPy wheel

License

Apache License 2.0 — see LICENSE for details.