llama.cpp's Model-Saver learns the SWA pattern and 15 more architectures now round-trip
GitHub·low signal
PR #29042, merged 2026-09-18, writes the sliding-window-attention pattern into saved models, which brings 15 additional architectures to a clean save-and-reload round trip. The same day's #28149 adds allocation-failure checks across ggml to turn out-of-memory crashes into errors, and #26070 handles graph buffer reservation failure. Three merges on one day all converting silent memory failures into reported ones.