Meta researcher's training trick: drop a batch proportion instead of per-sample to keep GPU efficiency

TimDarcet · x · 2026-09-09

Meta researcher TimDarcet shares a practical training trick: instead of dropping each sample with probability p, drop a proportion p of each batch. This keeps tensor shapes constant and preserves GPU efficiency. He says this is how it's done in Meta's open-source capi vision training library (see model.py), and notes that if everything stays on GPU it compiles through perfectly.

Related event: Meta researcher shares batch-ratio token dropping training trick(3 posts)→

Original post →

More from Research

Research channel →