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()) and pass generator=config.torch_rng() in eager mode; they automatically switch to the global RNG when compiled.