llama.cpp b11132 Adds Support for Gemma 4 DSpark Draft Models
llama.cpp release b11132 introduces GGUF format conversion and runtime inference support for Gemma 4 DSpark draft backbones. The update supports both full-attention and Sliding Window Attention (SWA) draft variants along with tied output weights and metadata-based feature inference. DSpark draft models enable speculative decoding to significantly accelerate LLM token generation speed. Bringing DSpark and Sliding Window Attention support for Gemma 4 into llama.cpp allows developers to run high-throughput inference more efficiently on consumer edge devices. The release adds boolean backbone metadata parameters and enables automatic feature detection via dflash. It also accounts for tied output weights across full-attention and SWA Gemma 4 draft architectures.
## BACKGROUND
Speculative decoding uses a smaller draft model to propose multiple candidate tokens that the main LLM verifies in parallel, speeding up generation. DSpark is a speculative decoding method designed to generate blocks of draft tokens efficiently. Sliding Window Attention (SWA) limits attention to a fixed recent context window, reducing KV cache growth and computational complexity from quadratic to linear.