PRNG

Where a datarax component's randomness comes from.

A component built from a seeded configuration receives an nnx.Rngs over the DEFAULT_RNG_STREAMS (augment, dropout, params, shuffling, default), built with substrax.rng.rngs_from_seed: each stream's key is derived from the seed and the stream's name, so a stream's key does not depend on the other streams present.

from substrax.rng import rngs_from_seed

from datarax.core.prng import DEFAULT_RNG_STREAMS

rngs = rngs_from_seed(42, DEFAULT_RNG_STREAMS)
key = rngs.augment()  # a raw key for an external library

A stochastic operator draws one key per record with per_record_keys, so a record's randomness depends only on the operator's base key, the epoch and the record's stable index, never on batch size, batch position or shuffle order.

See Also

• Operator - Where per-record keys are consumed
• Config - Seeding a component from configuration
• NNX Best Practices - PRNG patterns

---

::: datarax.core.prng