The ggml-org team released llama.cpp b11972 with fixes for two cases that could write beyond allocated buffers. One rejects control-vector layer indices above the model’s layer count; the other changes image-position loops in several vision encoders to stay within the patch count when image grids are not aligned to the merge size. PaddleOCR now also requires merge-aligned images when building its graph.
Why it matters
People running models with vision inputs or control vectors may avoid memory corruption in these cases, though the release notes do not quantify how often they occur.
