Installation

pip install rocci            # core: numpy only
pip install 'rocci[plot]'    # optional plotting support

Binary wheels ship for Linux (x86-64, aarch64, musl), macOS (Intel and Apple silicon), and Windows, for every Python ≥ 3.10. This means that a pip install rocci never needs to compile the rust components on a supported platform.

Supported install methods

=== "pip"

```bash
pip install 'rocci[plot]'
pip install rocci==0.2.0        # pin for reproducible environments
```

=== "uv"

```bash
uv add rocci                    # core: numpy only
uv pip install 'rocci[plot]'    # optional plotting support
```

=== "conda"

```bash
conda activate myenv
pip install rocci
```

rocci is not on conda-forge at present. The usual reason
to want a conda build is to coordinate shared native libraries across
packages. That does not apply here. The wheel is lightweight and
self-contained (the Rust kernel is statically linked; the only
dependency is numpy). Calling pip-inside-conda should suffice
 for most users. If a conda-forge package would matter for your setup,
 open an issue — demand is what would change the decision.

=== "from source"

```bash
# needs a Rust toolchain: https://rustup.rs
pip install git+https://github.com/ndelaneybusch/rocci
```

Installing from a git ref (or from the sdist on a platform with no
wheel) compiles the Rust kernel locally via maturin. Build time is a
couple of minutes.

The fallback backend

rocci's bootstrap kernel is compiled Rust. On a platform with no wheel and no Rust toolchain, the package still works: a pure-NumPy kernel with identical statistical semantics takes over automatically, and rocci emits a FallbackBackendWarning once per process so you know you are on the slow path (10–30× slower). Which backend is active is recorded on every result as band.backend and printed by rocci.show_versions().

There is nothing to configure: backend selection is automatic and does not change any statistical output contract. Same seed + same backend + same version ⇒ bit-identical bands; across backends, results agree statistically but not bit-for-bit (different RNG streams — documented, tested).

Requirements

Python	≥ 3.10
Hard dependencies	numpy >= 1.24
Optional	matplotlib >= 3.7 (rocci[plot]), pandas (only for to_dataframe())

For bug reports

import rocci
rocci.show_versions()

prints the version, active backend, dependency versions, OS, and CPU count — paste it into the issue template.