⚙️🎲 HAlea

Unified access to (some) hardware RNGs for machine learning & scientific computing

Supported Devices

Features

• Three APIs: NumPy (HaleaNumpy), PyTorch (HaleaTorch), JAX (HaleaJax)
• Distributions: uniform, normal (Box-Muller), integers, bits, bytes
• Sampling: choice() with/without replacement, shuffle(), permutation()
• Auto-detection: Finds connected devices automatically
• Buffered I/O: Efficient bulk reads with configurable buffer size
• Unbiased mode: Optional rejection sampling for perfect uniformity

Quick Start

from halea import HaleaNumpy

with HaleaNumpy() as rng:
    # Uniform [0, 1)
    x = rng.rand(1000)

    # Normal N(0, 1)
    y = rng.randn(100, 100)

    # Integers [0, 100)
    z = rng.randint(100, size=50)

    # Shuffle in-place
    rng.shuffle(my_list)

PyTorch / JAX:

from halea import HaleaTorch, HaleaJax

with HaleaTorch(device="cuda") as rng:
    tensor = rng.randn(256, 256, dtype=torch.float32)

with HaleaJax() as rng:
    array = rng.normal(shape=(256, 256))

Installation

pip install halea

Requirements: numpy, pyserial, pyusb Optional: torch, jax (for respective APIs)

Linux udev rules (for non-root access):

# TrueRNG
echo 'SUBSYSTEM=="tty", ATTRS{idVendor}=="04d8", ATTRS{idProduct}=="f5fe", MODE="0666"' | sudo tee /etc/udev/rules.d/99-truerng.rules

# ChaosKey
echo 'SUBSYSTEM=="usb", ATTR{idVendor}=="1d50", ATTR{idProduct}=="60c6", MODE="0666"' | sudo tee /etc/udev/rules.d/99-chaoskey.rules

sudo udevadm control --reload-rules

Scope & Limitations

• For ML/scientific use — not cryptographic applications
• Single-threaded — create separate instances for multi-threaded use
• No entropy monitoring — assumes device is functioning correctly
• Linux-focused — other platforms may work but are untested

Security Notice

This library provides access to hardware random number generators for machine learning and scientific computing applications. It is NOT designed for cryptographic use.

For cryptographic applications:

1. The default randint() uses float-based scaling with negligible bias (~1e-15) for ranges under 2^53. Use unbiased=True for statistically perfect uniformity.

2. No entropy health monitoring is performed. If a device fails, the library may return low-quality randomness.

3. For cryptographic random bytes, use Python's secrets module or read from /dev/random directly.