nemo_automodel.components.speculative.dflash

View as Markdown

DFlash speculative-decoding training components.

DFlash drafts a whole block of tokens in parallel via MASK-token “denoising” conditioned on the target model’s hidden states, in contrast to EAGLE’s autoregressive single-step drafting. See nemo_automodel.components.speculative.dflash.core for the training wrapper.

Submodules

Package Contents

Data

__all__

API

nemo_automodel.components.speculative.dflash.__all__ = ['DFlashTrainerModule', 'DFlashStepMetrics', 'NoValidAnchorsError', 'create_dfla...