Announcement
@jadenfk23 announces NNterp 2.0 to tame transformer-model quirks
NNterp 2.0 lets one NNsight experiment span hundreds of transformer models, according to @jadenfk23.
TLDR
Write once, swap models: @jadenfk23 says NNterp 2.0 lets the same NNsight experiment run across hundreds of transformers. @davidbau describes a unified API for accessing components such as decoder heads, residual streams and attention, with support for linear attention, state-space models, mixture-of-experts and vision encoders. Less model-by-model plumbing is the pitch.
Combined views
175
5 Sources, first seen ago

Comments