llama.cpp Release b10231 Introduces Support for DSpark Speculative Sidecars
The llama.cpp release b10231 introduces support for DSpark speculative sidecar resolution. This update allows dspark- files to resolve similarly to other speculative sidecars, prioritizing DSpark over DFlash in auto-selection due to its inclusion of an extra Markov head. Speculative decoding is crucial for accelerating LLM inference, and integrating DSpark support enables more efficient draft generation and verification. This helps developers achieve higher throughput and lower latency when running compatible models on llama.cpp. The update applies the -hfd tag to DSpark files, allowing a requested sidecar to resolve without requiring a full model at the tag, while an explicit -md selection disables this discovery. If no specific type is requested, DSpark is auto-selected over DFlash because its sidecar carries an extra Markov head.
## BACKGROUND
Speculative decoding accelerates LLM inference by using a smaller, faster draft model (or "sidecar") to propose tokens, which are then verified in parallel by the larger target model. DSpark is a speculative decoding framework developed by DeepSeek that utilizes a Markov head to improve draft generation and verification efficiency. llama.cpp is a popular open-source C/C++ inference engine optimized for running LLMs on various hardware platforms.