hard20 pts · first answer countsreproducibilitygpu

Same seed, different results

You set the same random seed, code and data, yet two GPU training runs give slightly different final metrics. What is the most likely reason?

A.The seed is ignored by all ML frameworksB.The data was secretly changed by the GPUC.Some GPU kernels are non-deterministic (for example, atomic adds change floating-point summation order), and data loader workers or library versions can also differD.Metrics are random by definition

Discussion 0

Join free to post in the discussion.

Same seed, different results

Discussion 0

Related questions