Randomness

std.random draws the same random numbers on every backend from keys you pass; there is no hidden generator, so the same key reproduces a run. A key is a Tensor[2; i64] of two 32-bit words, expanded with Threefry-2x32.

use std.random::{categorical, fold_in, normal, split, uniform}

let keys = split<Steps>(key)          // one derived key per step
let noise = normal<N, f32>(keys[0, :])
let u = uniform<N, bf16>(fold_in(key, 3))
let token = categorical(key, logits)  // one draw per row, proportional to exp(logits)

Function	Result
threefry2x32<N>(key, counter: Tensor[N, 2; i64])	N output pairs
bits<N>(key)	N random 32-bit words as i64
split<N>(key)	N new keys
fold_in(key, data: i32)	a key derived from key and an integer
uniform<N, T>(key)	N numbers in [0, 1)
normal<N, T>(key)	N standard normal numbers (Box-Muller)
categorical<B, V, T>(key, logits)	one category per row (Gumbel-max)

Agreement with JAX

For a key from jax.random.key(seed) (key_data gives the two words), split, bits, uniform, and fold_in match jax.random exactly under its default partitionable Threefry layout. normal and categorical agree across Linnet backends but not with JAX.