hyperfold

High-performance linear algebra without leaving Python.

A modern C++23 computing core with a NumPy-first API, explicit memory behaviour, and native CPU and CUDA execution.

PYPython API (NumPy, CuPy and PyTorch)
APIZero-copy CuPy and PyTorch interchange
SIMDNative and extensible CPU kernels (natively supporting AVX2, AVX-512 and Apple Silicon NEON)
CUDANative and extensible CUDA kernels (natively supporting up to the Blackwell architecture)
LINALGDense and structured linear algebra
SOLVERSolver and factorisation pipelines

import cupy as cp
import torch
import hyperfold as hf

H = [[4, 1-1j], [1+1j, 3]]  # Hermitian
A = cp.asarray(H, dtype=cp.complex64)
b = cp.asarray([1+2j, 3-1j], A.dtype)

lu = hf.lu_factor(A)       # factor once
x1 = lu.solve(b)
x2 = lu.solve(2 * b)       # solve many

t = torch.ones(8, device="cuda")
y = hf.scale(t, 2)          # zero-copy

braketware

hyperfold