gigl.common.utils.compute.random#

Matches the set_seed(seed, deterministic=False) shape used by Hugging Face Transformers, MMEngine, and Accelerate; follows the recipe at https://pytorch.org/docs/stable/notes/randomness.html.

Attributes#

logger

Functions#

seed_everything([seed, ...])

Seed Python / NumPy / PyTorch RNGs, optionally enforce deterministic torch ops.

Module Contents#

gigl.common.utils.compute.random.seed_everything(seed=_DEFAULT_SEED, should_enable_expensive_deterministic_compute=False)[source]#

Seed Python / NumPy / PyTorch RNGs, optionally enforce deterministic torch ops.

What gets seeded:

  • random.seed(seed) — Python stdlib.

  • np.random.seed(seed) — NumPy global RNG.

  • torch.manual_seed(seed) — CPU and all CUDA devices (torch.manual_seed calls torch.cuda.manual_seed_all internally. Also covers PyTorch Geometric.

When should_enable_expensive_deterministic_compute=True (opt-in; default False because it costs throughput and should not be enabled for training or for production inference - can be used for debugging purposes.

  • Important: Graph Sampling currently do not follow determinism outlined here.

Example

>>> seed_everything(42)
42
Parameters:

  • seed (int) – RNG seed.

  • deterministic – If True, also enforces bitwise-deterministic torch ops (cudnn flags, use_deterministic_algorithms, CUBLAS_WORKSPACE_CONFIG). Default False — most training pipelines want seeded RNGs without paying the throughput cost.

  • should_enable_expensive_deterministic_compute (bool)

Return type:

None

gigl.common.utils.compute.random.logger[source]#