Implementation:AUTOMATIC1111 Stable diffusion webui Philox RNG

Knowledge Sources	AUTOMATIC1111_Stable_diffusion_webui
Domains	Random_Number_Generation, Reproducibility
Last Updated	2025-05-15 00:00 GMT

Overview

Implements the Philox 4x32 counter-based random number generator with Box-Muller transform on CPU, replicating the exact output of torch.randn on CUDA devices for cross-device reproducibility.

Description

The Philox RNG module provides a pure NumPy implementation of the Philox 4x32 counter-based random number generator, matching the CUDA implementation used by PyTorch. The philox4_32 function runs 10 rounds of the Philox algorithm on a 4xN counter array and a 2xN key array (derived from the seed), producing 4xN pseudo-random 32-bit integers. The box_muller function transforms pairs of uniform random integers into standard normal random variables using the Box-Muller method, returning only the sine component. The Generator class provides a user-friendly interface: given a seed, it generates normally distributed random arrays of arbitrary shape via the randn method, maintaining an internal offset counter that increments with each call. This allows CPU-based generation that produces identical results to torch.randn(..., device='cuda'), enabling reproducible image generation regardless of the device used.

Usage

Use this module when you need to generate random noise tensors on CPU that exactly match the output of CUDA's random number generator, ensuring that the same seed produces the same image regardless of whether generation runs on CPU or GPU.

Code Reference

Source Location

• Repository: AUTOMATIC1111_Stable_diffusion_webui
• File: modules/rng_philox.py
• Lines: 1-102

Signature

def uint32(x: np.ndarray) -> np.ndarray
def philox4_round(counter: np.ndarray, key: np.ndarray) -> None
def philox4_32(counter: np.ndarray, key: np.ndarray, rounds: int = 10) -> np.ndarray
def box_muller(x: np.ndarray, y: np.ndarray) -> np.ndarray

class Generator:
    def __init__(self, seed: int) -> None
    def randn(self, shape: tuple) -> np.ndarray

Import

from modules.rng_philox import Generator

I/O Contract

Inputs

Name	Type	Required	Description
seed	int	Yes	The random seed, used as the Philox key
shape	tuple[int, ...]	Yes	Shape of the output random array
counter	np.ndarray	Yes	4xN array of uint32 counter values (internal use)
key	np.ndarray	Yes	2xN array of uint32 key values derived from the seed (internal use)
rounds	int	No	Number of Philox rounds (default 10)

Outputs

Name	Type	Description
result	np.ndarray	Array of standard normal random values with the requested shape (float32)

Usage Examples

from modules.rng_philox import Generator

g = Generator(seed=42)

# Generate noise matching torch.randn output on CUDA
noise = g.randn(shape=(1, 4, 64, 64))

# Subsequent calls advance the internal offset
noise2 = g.randn(shape=(1, 4, 64, 64))

Related Pages

• Principle:AUTOMATIC1111_Stable_diffusion_webui_Seed_Reproducibility