bridge.training.post_training.dspark

DSpark speculative-decoding draft training (architecture-agnostic core).

This package holds the pure-torch pieces of DSpark draft training: the sequential (Markov / RNN) and confidence heads (:mod:heads) and the three-term CE / L1 / confidence objective with its acceptance metrics (:mod:loss). The parallel-aware draft backbone and the Megatron-Core forward-step / provider wiring are separate, GPU-validated concerns. See docs/training/dspark-speculative-decoding.md.

Submodules

• bridge.training.post_training.dspark.loss
• bridge.training.post_training.dspark.heads

Package Contents

Data

__all__

API

bridge.training.post_training.dspark.__all__

[‘ConfidenceHead’, ‘DSparkForwardOutput’, ‘GatedMarkovHead’, ‘RNNHead’, ‘VanillaMarkov’, ‘assert_per…