This document is relevant for: Trn2, Trn3

Speculation

Design documentation for speculative decoding. For configuration guidance, see Speculative decoding in the features guide and the EAGLE3 tutorials for Llama 3.1 and GPT-OSS.

Topic	Description
Speculative decoding (EAGLE3)	EAGLE3 autoregressive drafter internals on Neuron
Speculative decoding (DFlash)	DFlash block-parallel drafter internals on Neuron

This document is relevant for: Trn2, Trn3