TL;DR: Build b11633 patches a CPU-backend bug where extra threads wrote past the work buffer during F16→F32 row conversion, preventing memory corruption in local inference.
Summary: llama.cpp release b11633 (PR #30454) fixes a ggml-cpu bug where the SET_ROWS operation was planned with a single task but executed by every thread, each using its own row of the F16→F32 conversion buffer. When the compute graph also contained another multi-threaded op, threads beyond the first wrote past the allocated work buffer, corrupting memory. The fix moves set_rows into an n_task = n_threads group so the conversion buffer is correctly sized for all threads.
Why it matters: Anyone running llama.cpp CPU inference on multi-core machines was exposed to silent memory corruption that could cause crashes or wrong outputs. Update to build b11633 or later, especially for Apple/mobile builds that bundle the XCFramework.
Source: github_releases