Skip to content

determinism.md

Latest commit

 

History

History
31 lines (22 loc) · 1.16 KB

determinism.md

File metadata and controls

31 lines (22 loc) · 1.16 KB
Raw

Determinism

Reproducibility requires the same configuration, device, software build, and CPU thread count. A seed alone does not guarantee identical results across systems.

experiments/runners/common.py sets:

  • CUBLAS_WORKSPACE_CONFIG=:4096:8 before CUDA initialization, unless already set.
  • Deterministic PyTorch algorithms with warn_only=True.
  • Deterministic cuDNN behavior, with benchmarking and TF32 disabled.
  • Seeded DataLoader generators and worker initialization.

Warnings remain possible when an operation has no deterministic implementation. Use torch.use_deterministic_algorithms(True, warn_only=False) to make these errors explicit when validating a workload.

Rotations

The rotation sampler uses batched QR for small particle batches and Householder reflections for larger batches. Both sample Haar rotations but consume different random streams. Changing the particle count or device can change the sequence.

Checks

pytest tests/test_determinism.py
pytest tests/test_determinism.py -m gpu

Checkpoint continuation also requires the original optimizer configuration and model weights; see the API reference.