Meta presents Layer Skip.

Enabling early exit inference and self-speculative decoding.

We present LayerSkip, an end-to-end solution to speed-up inference of large language models (LLMs).

---

Join the discussion on this paper page.