Generator

Reproducible seeding and state management. The Generator wraps torch.Generator for CPU/CUDA today; on Neuron hardware it will dispatch to a Philox counter-based RNG running on the GpSimd engine.

Generator(seed=None, device="cpu")

Construct a seeded generator. seed=None leaves the underlying torch.Generator unseeded (random state).

Generator.manual_seed(seed) -> Generator

Re-seed in place. Returns self so you can chain.

Generator.get_state() / set_state(state)

Save and restore the underlying torch.Generator state for checkpointing.

Generator.torch_generator

Access the underlying torch.Generator for use with raw PyTorch ops.

manual_seed(seed) -> Generator

Module-level helper that re-seeds (and replaces) the default generator.

get_default_generator() -> Generator

Return the module-level default generator used when generator=None is passed to a distribution function.