Sentence Transformers v6.0: fp16 + FlashAttention Hits 3.87x Over fp32

tomaarsen · x · 2026-08-18

v6.0 performance notes: multi-column losses now run a single forward pass over merged columns, 1.25x faster on hard-negative & triplet training with identical loss trajectories; fp16 + FlashAttention is now the fastest GPU config at 3.87x over fp32. Breaking changes include transformers v5 & torch 2.2 as the floor, similarity/similaritypairwise becoming methods, trustremotecode=True required for custom module classes (including local paths), and a bare chat message list meaning one conversation instead of a batch.

Related event: Sentence Transformers v6.0 Ships Late-Interaction Multi-Vector Models(27 posts)→

Original post →

More from coding & agent

coding & agent channel →