ggerganov/llama.cpp Released Build b11907
Release b11907 of llama.cpp replaces ggml_tensor output pointer parameters with dedicated output structures for graph builders. This change refactors how extra outputs such as gates and top-k selections are handled in hyper-connection and DSA graph builders. This refactoring improves code cleanliness, type safety, and maintainability for internal graph builders within the GGML machine learning ecosystem. Developers working closely with custom graph operations or model architectures will need to update their implementations to align with the new output structure pattern. The update specifically touches hyper-connection and DSA graph builders, returning extra outputs alongside primary outputs in a small struct rather than using raw ggml_tensor** pointers. The release also includes numerous precompiled binaries across major platforms including macOS, Linux, Windows, and Android.
## BACKGROUND
llama.cpp is an open-source software library that enables efficient inference for large language models on various hardware using the GGML tensor library. Graph builders in this context construct computational graphs to execute neural network operations layer by layer.