Install¶

Python¶

pip install perturbation-kernel

Wheels ship for Linux, macOS and Windows. They are built against the stable ABI, so one wheel per platform serves every CPython from 3.8 onward and you do not need a Rust toolchain.

The GPU backend is compiled into the published wheels. wgpu loads its driver lazily and reports no adapter on a machine without one, so the wheel installs and runs identically on a headless box. Check what you actually got:

import perturbation_kernel as pk

pk.available_backends()   # ['auto', 'scalar', 'simd', 'gpu']
pk.simd_path()            # 'neon' | 'avx2' | 'scalar'
pk.gpu_device()           # 'Apple M4 Max (Metal, IntegratedGpu)' or None

From source¶

git clone https://github.com/godofecht/perturbation-kernel
cd perturbation-kernel/python
pip install maturin
maturin develop --release

Rust¶

[dependencies]
perturbation-kernel = "2"

Requires Rust 1.85 or later, or 1.87 with the gpu feature, which depends on wgpu. Default features are parallel and simd, both of which are bit-identical to the reference path.

# Portable scalar only: no rayon, no intrinsics.
perturbation-kernel = { version = "2", default-features = false }

# With the wgpu compute backend.
perturbation-kernel = { version = "2", features = ["gpu"] }

Feature	Default	Effect
`parallel`	yes	Spreads the draw loop across a `rayon` pool above 4096 draws
`simd`	yes	NEON on aarch64, AVX2 on x86-64, for the reductions
`gpu`	no	Adds `wgpu`, enabling `Backend::Gpu` for the built-in families

Disabling either default feature changes how long a run takes and nothing else. There is no configuration under which the numbers move.

C and C++¶

The crate builds a cdylib and a staticlib carrying the C ABI:

cargo build --release
# target/release/libperturbation_kernel.{dylib,so,a}

See the C API reference.

Lean 4¶

cd lean/PerturbationKernel
lake build

Verifying the install¶

The fastest meaningful check is that two backends agree exactly:

import perturbation_kernel as pk

a = pk.Markov(k=5, theta_max=0.3).run(pk.Config(n=50_000, seed=7))
b = pk.Markov(k=5, theta_max=0.3).run(pk.Config(n=50_000, seed=7, backend="scalar"))
assert a.value.hex() == b.value.hex()

If that fails, the build is broken in a way that matters. Please open an issue with the output of pk.simd_path() and your platform.