DeepSeek D-Spark

Mixes the drafting advances of DFlash (parallel) with EAGLE-3/MTP (sequential).

TODO: Parallel outputs ? logits?

Sequential is a lightweight Markov model.

The ‘optimal’ number of draft tokens depends on the domain. There is high domain variance as code is relatively ‘predictable’ (low-entropy) and therefore you can draft further ahead, but chat is more open ended (high-entropy).

Novelty: confidence head outputs how likely a token is to be accepted/rejected. This means we can draft N tokens and verify the first M (M<=N). They also make this process hardware aware, so based on utilisation it can ensure it is optimised to do more or less.

Links:

• How Speculative Decoding Makes LLMs Faster Without Retraining (and What DSpark Adds)
• DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
• DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence - Huggingface
• DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence - Paper