The ggml‑org/llama.cpp library released a new version on Oct 21 2026. It changes the LFM2 image processor to keep the original aspect ratio and to use the same rounding logic as the Python code used by transformers. With the new PAD_NONE setting and double‑precision rounding, tile sizes now match those produced by transformer models for every image size up to 4096 × 4096. The fix removes black bars and inconsistent tile grids that previously affected d1‑3B and other LFM2‑VL models.
Why it matters
Developers using LFM2 vision models will see more reliable image preprocessing, avoiding size‑related bugs when swapping between llama.cpp and transformer code.
