Two researchers run the same model with the same random seed and obtain slightly different results. That does not automatically mean one copied the seed incorrectly. A seed controls a particular source of pseudorandomness; it does not freeze every operation, library version or device behaviour in an experiment.
PyTorch’s reproducibility guidance explicitly warns that identical results are not guaranteed across releases, platforms or CPU and GPU execution, even with the same seed. It distinguishes controlling random sources from selecting deterministic operations.
Define what you want to repeat
Repeating a run on the same machine with the same environment is a narrower goal than reproducing it after a framework upgrade on different hardware. State that scope in the experiment record. Otherwise, two people may use the word reproducible while testing different promises.
Record the code revision, data version and split, dependencies, device, relevant settings and seed values. Include preprocessing and data-loader behaviour. A model configuration alone does not describe all the work that produced its inputs.
Separate debugging from evaluation
For debugging, a repeatable run helps isolate whether a code change caused an observed difference. For evaluating a model, repeating only one favourable seed can hide variability. These are different uses of controlled randomness.
As a simple experiment plan, first repeat an unchanged configuration in the same environment. Then change one intended factor while preserving the rest of the record. If results differ unexpectedly, compare the saved environment and data information before inventing an explanation about model quality.
Deterministic settings can affect speed or operation availability, so consult the framework’s current documentation rather than copying a universal setup block from an unrelated project. Record any trade-off that matters to the experiment.
When publishing a result, distinguish “we obtained this metric in this run” from “this configuration reliably improves the metric.” The latter needs evidence about repeated evaluation and variation. A seed is useful evidence about how a run began, but it is only one part of the record needed to understand how the run unfolded.