Speculative decoding

A draft model proposes tokens that the target model verifies. For SpeechLM, DFlash and DFlash2 drafts use intermediate audio-conditioned target features. Start with inference using a compatible exported draft, then see training to adapt a draft to a different target checkpoint.

Inference

• DFlash speculative decoding with vLLM
  • Requirements
  • DFlash2 inference
  • Validation

Training

• DFlash and DFlash2 draft training
  • Launching training