DeepSeek's DSpark Makes Speculative Decoding Practical for Production LLMs

DeepSeek's DSpark paper introduces a method to graft a speculative decoding head directly onto a target model, avoiding the need for a separate draft model. This reduces layer duplication and can boost throughput 2-4x while maintaining lossless output. The technique is complementary to Multi-Token Prediction and is open-sourced in the DeepSpec repository.