The llama.cpp project released build b10052, featuring significant internal changes to the Hexagon backend. The update includes a rework of L2 cache handling with dirty bit tracking and lazy flushing, alongside various updates to matrix multiplication kernels.
- hex-mm: fixes an artificial limit in the solver that restricted the number of act-prep threads
- hmx-mm: adds support for tiled act-processing to better distribute Hvx work
- hex-l2: optimizes L2 flushes at the opbatch level and tracks tensor aliases to delay or skip flushes
- hex-work: redoes the legacy workpool API to match hmx-queue and dma-queue
- hex-main: introduces a fallback mode for dspqueue callbacks and removes early wakeup logic that caused errors
This release provides updated binaries for macOS, Linux, Windows, Android, and iOS across CPU, GPU, and specialized hardware backends.