pytorch config_error ai_generated true

Results not reproducible: different outputs across runs

ID: pytorch/rng-reproducibility

Also available as: JSON · Markdown
78%Fix Rate
82%Confidence
3Evidence
2023-01-01First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
2 active — — —

Root Cause

Model producing different results across runs.

generic

Workarounds

  1. 88% success Set all seeds: torch.manual_seed, np.random.seed, random.seed
  2. 90% success Use torch.use_deterministic_algorithms(True) and set CUBLAS_WORKSPACE_CONFIG

Dead Ends

Common approaches that don't work:

  1. Use deterministic mode for everything 72% fail

    Performance degradation on some ops

  2. Ignore reproducibility 75% fail

    Cant debug or verify results