nemo_rl.models.automodel.draft#

DSpark/DFlash/EAGLE3 draft co-training on the Automodel backend.

Thin extension layer over nemo_automodel.components.speculative (3rdparty Automodel r0.6.0): byte-identical building blocks are imported from Automodel; only the RL-co-training deltas live in this package (see each module’s header for what it extends and why). The draft model predicts a block of tokens per anchor position in one parallel pass, cross-attending target-model hidden states captured from the policy’s own training forward.

Submodules#

Package Contents#

Data#

__all__

API#

nemo_rl.models.automodel.draft.__all__#

[‘DSparkForwardOutput’, ‘Qwen3DSparkModel’, ‘build_draft_config’, ‘compute_dspark_loss’, ‘create_dfl…