The ggml‑org/llama.cpp project released version b11633 on Oct 15 2026. The update changes how the SET_ROWS operation allocates its conversion buffer, ensuring each thread gets its own space. Previously, threads beyond the first wrote past the buffer, causing crashes in multi‑threaded runs. The fix moves SET_ROWS to run with one task per thread, aligning buffer size with the thread count.
Why it matters
Developers using llama.cpp on multi‑core CPUs will see fewer crashes when running models in parallel.
