get_rng_state

→Tensor

get_rng_state()

source edit

Snapshot the full state of the default random generator.

Returns an int64 tensor of length 2, [seed, counter], that completely characterises the default Philox generator.

The Philox-4×32 PRNG is a stateless counter-based design: the entire infinite sample stream is determined by the two-tuple

\text{state} = (k,\; c) \quad k \in [0,\, 2^{63}),\; c \in [0,\, 2^{64})

where $k$ is the seed key and $c$ is the call counter. After each draw of a 128-bit output block, $c$ is incremented by 1. The next sample is always $\text{Philox}(k, c)$ , so saving $(k, c)$ is a lossless checkpoint of the generator.

Returns

Tensor

1-D int64 tensor [seed, counter] of shape (2,). Pass this directly to set_rng_state to restore the stream.

Notes

Unlike Mersenne Twister (which requires 624 × 32-bit words of state), Philox's full state is just 128 bits. This makes checkpointing essentially free and enables exact reproducibility of training runs when saved alongside model weights.

Examples

>>> import lucid
>>> lucid.manual_seed(0)
>>> _ = lucid.rand(10)           # advance counter by drawing 10 values
>>> state = lucid.get_rng_state()
>>> a = lucid.rand(5)
>>> lucid.set_rng_state(state)   # rewind
>>> b = lucid.rand(5)
>>> (a - b).abs().max().item()
0.0

>>> import lucid >>> lucid.manual_seed(0) >>> _ = lucid.rand(10) # advance counter by drawing 10 values >>> state = lucid.get_rng_state() >>> a = lucid.rand(5) >>> lucid.set_rng_state(state) # rewind >>> b = lucid.rand(5) >>> (a - b).abs().max().item() 0.0