I am running the same program with the same input dataset and configuration, but I sometimes get different results each time I execute it. The differences are not minor rounding errors — the model performance or output can noticeably vary between runs.

I have already confirmed that:

• The input data has not changed

• The code has not been modified

• The execution environment appears to be the same

However, the outputs are still inconsistent.

I would like to understand:

• What are the most common reasons why identical code can produce different results across runs?

• How can randomness in training processes (e.g., initialization, data shuffling, dropout, or hardware-level operations) affect reproducibility?

• Can GPU computation or multi-threading introduce non-deterministic behavior even when everything seems fixed?

• What are the best practices to ensure reproducible results in machine learning or data processing pipelines?