The ggml-org/llama.cpp b12030 release fixes a Hexagon memory-transfer bug affecting strides of 16 MB or more. On Hexagon v75 and later, the DMA descriptor holds row sizes and strides in 24 bits; the old fallback split transfers by row count but could still truncate oversized strides, leading to incorrect addresses. The fix uses one transfer per row in that case, and the release adds two CONT tests with source strides above 16 MB.
Why it matters
For people running llama.cpp on affected Hexagon hardware, the fix should prevent incorrect memory addresses during transfers with very large strides; the release notes do not say which workloads encounter them.
