Llama.cpp Adds Support for DeepSeek's DSpark Speculative Decoding

A new pull request (#25173) in the llama.cpp repository adds support for DSpark speculative decoding, a framework developed by DeepSeek. This integration allows users to run models like DeepSeek-V4 with DSpark-enabled checkpoints for faster local inference. Speculative decoding significantly improves LLM inference speeds without losing accuracy, and integrating DSpark into llama.cpp makes high-performance local AI inference more accessible. This optimization is particularly beneficial for running large models like DeepSeek-V4 on consumer hardware. DSpark is an inference optimization rather than a new model, reusing existing DeepSeek-V4 weights with an attached draft module to accelerate generation. The pull request has prompted community members to begin benchmarking performance improvements, specifically tracking prompt processing (pp) and token generation (tg) speeds.