set_rng_state

→None

set_rng_state(state: Tensor)

source edit

Restore the default random generator to a previously captured state.

Rewinds (or fast-forwards) the Philox generator to the exact (seed, counter) pair encoded in state, so that the next sample call produces exactly the same value as it did immediately after the get_rng_state call that created state.

Formally, after set_rng_state(state) the generator satisfies:

k \leftarrow \texttt{state}[0], \quad c \leftarrow \texttt{state}[1]

and the next draw returns $\text{Philox}(k, c)$ .

Parameters

stateTensor

Length-2 int64 tensor [seed, counter] as returned by get_rng_state. Passing a tensor of any other shape or dtype raises ValueError.

Raises

ValueError

If state does not have exactly 2 elements.

Notes

set_rng_state + get_rng_state together provide a lightweight fork–join pattern for reproducible stochastic computation:

s^* = \text{get\_rng\_state}() \;\to\; \text{sample branch A} \;\to\; \text{set\_rng\_state}(s^*) \;\to\; \text{sample branch B}

Both branches then draw from the same sub-stream, enabling controlled ablations where only the model changes, not the data augmentation noise.

Examples

>>> import lucid
>>> lucid.manual_seed(99)
>>> state = lucid.get_rng_state()
>>> x = lucid.randn(4)
>>> lucid.set_rng_state(state)
>>> y = lucid.randn(4)
>>> (x - y).abs().max().item()
0.0