nemo_automodel.components.speculative.dspark
nemo_automodel.components.speculative.dspark
DSpark speculative-decoding draft model and training objective.
A semi-autoregressive parallel drafter: a parallel backbone produces every position of a block in one pass, a lightweight serial Markov head injects intra-block token dependency, and a confidence head predicts per-position acceptance probability for scheduled verification.
Submodules
nemo_automodel.components.speculative.dspark._samplingnemo_automodel.components.speculative.dspark.commonnemo_automodel.components.speculative.dspark.confignemo_automodel.components.speculative.dspark.corenemo_automodel.components.speculative.dspark.draft_gemma4nemo_automodel.components.speculative.dspark.draft_qwen3nemo_automodel.components.speculative.dspark.lossnemo_automodel.components.speculative.dspark.markov_headnemo_automodel.components.speculative.dspark.registrynemo_automodel.components.speculative.dspark.target