The llama.cpp project released build b11039, which updates the model-saver to write sliding window attention (SWA) patterns and MLA SWA geometry for all models using SWA. The change ensures that the `sliding_window_pattern` is written as a per-layer flag rather than being collapsed to a scalar, preserving exact data representation.

  • Loaders now read the SWA pattern as either a period or a per-layer array via `llama_model_base::load_swa_pattern()`, fixing silent ignores of arrays in converters for olmo2, gemma3n, and exaone4.
  • The model-saver writes MLA key/value lengths and KV LoRA rank for SWA layers, which is required by dots3note.
  • This enables the saver for plamo3, gemma3, cohere2, cohere2moe, olmo2, exone-moe, afmoe, mimo2, spark2_5, muse-glimmer, mellum, laguna, granite_swa, dots3note, and maple.
  • All listed models pass the bit-exact roundtrip of test-llama-archs.

The update ensures that GGUF files generated by llama.cpp match the intended SWA configurations without altering outputs for existing published files that relied on default periods.