llama.cpp b10679 adds lazy tensor read benchmark

TL;DR: llama.cpp's b10679 release adds a --tensor-read-lazy benchmark option to measure lazy tensor loading performance.

Summary: ggml-org released llama.cpp b10679, adding a --tensor-read-lazy flag to the benchmark tool (PR #27881). The change renames the feature to LLAMA_LAZY_MODE_* and removes the old alias, refining how lazy tensor reads are measured. Builds are available for Linux, Windows, macOS/iOS, and Android.

Why it matters: For AI builders, this improves visibility into lazy tensor loading, which can reduce memory overhead when running large models on constrained devices. Try the new flag in llama-bench to compare lazy vs. eager tensor read performance.

Source: github_releases