llama.cpp b10231 adds DSpark sidecar support for faster speculative decoding
DSpark draft models now auto-discover and outrank DFlash for accelerated inference...
Deep Dive
ggml-org's llama.cpp released b10231, adding DSpark speculative sidecar resolution. The update supports -hfd tag handling, allows a requested sidecar to resolve without a full model at the tag, and lets an explicit -md selection disable discovery. When no type is requested, DSpark outranks DFlash in auto-selection since its sidecar carries the extra Markov head.
Key Points
- DSpark sidecars now support -hfd tag resolution and auto-discovery without a full model
- Explicit -md selection disables automatic sidecar discovery; DSpark outranks DFlash by default due to its Markov head
- Prebuilt binaries available for macOS, Linux, Windows, Android, iOS, and openEuler across CPU, Vulkan, CUDA 12/13, ROCm 7.2, OpenVINO, and SYCL
Why It Matters
DSpark support makes speculative decoding more automatic and robust, cutting local LLM latency and resource costs for developers.