Random number generation#

In order to make your simulations reproducible, it is important to configure a random seed which makes your code deterministic. When Sionna is loaded, the Config singleton instantiates random number generators (RNGs) for Python, NumPy, and PyTorch. You can then set a single seed which will make all of your results deterministic, as long as only these RNGs are used. In the cell below, you can see how seed is set and how py_rng, np_rng, and torch_rng() can be used in your code. All of Sionna PHY’s built-in functions rely on these RNGs.

import torch
from sionna.phy import config
config.seed = 40

# Python RNG - use instead of
# import random
# random.randint(0, 10)
print(config.py_rng.randint(0, 10))

# NumPy RNG - use instead of
# import numpy as np
# np.random.randint(0, 10)
print(config.np_rng.integers(0, 10))

# PyTorch RNG - use instead of
# torch.randint(0, 10, (1,))
print(torch.randint(0, 10, (1,), generator=config.torch_rng(), device=config.device))

7
5
tensor([7])

For code that uses torch.compile, use the compile-aware utilities in the utility functions section of the PHY API (e.g. randint(), normal()). Their output defaults to config.device. When passing an explicit generator in eager mode, pass the matching device as well:

from sionna.phy.utils import normal

noise = normal(
    [4],
    device=config.device,
    generator=config.torch_rng(config.device),
)

The compile-aware helpers automatically switch to the seeded global RNG when compiled because explicit generators cannot be captured in the graph.

What config.seed guarantees#

Setting seed reinitializes Sionna’s configured Python, NumPy, and per-device PyTorch generators. Public stochastic behaviour in Sionna PHY and Sionna SYS that goes through those generators is then reproducible across runs that:

  • use the same seed,

  • use the same code path (eager vs compiled; see below),

  • and do not draw from other process-global RNGs (random, np.random, or unseeded torch.* calls without a generator).

Typical examples that follow this contract include AWGN / Rayleigh channel draws, discrete-channel bit flips, OFDM Kronecker pilot symbols, CIRDataset shuffle order, HexGrid UT placement, and FEC scramblers / interleavers that take a seed or use the configured generators.

What it does not guarantee#

  • Eager vs compiled equality. In eager mode the compile-aware helpers use config.torch_rng(device). Under torch.compile they fall back to the seeded global device RNG. Each mode is individually reproducible after config.seed = ..., but the two modes are not required to produce bitwise-identical samples for the same seed.

  • Multi-worker ``DataLoader``s. Sharing one config.py_rng across worker processes is not a supported reproducibility model. Keep num_workers=0 (the Sionna default for CIRDataset) or give each worker an explicit local generator.

  • Sionna RT preview cosmetics. Default radio-material colours and other preview-only randomness in the RT submodule are outside sionna.phy.config.seed. Electromagnetic sampling in RT uses Mitsuba samplers with their own explicit seeds.

  • Unrelated global Torch state. Code that calls torch.rand / torch.randn without passing generator=config.torch_rng(...) (or the compile-aware wrappers) is not controlled by config.seed in eager mode.