The ggml-org team’s llama.cpp b11913 release adds a configurable size override for VTCM, a memory resource used by the Hexagon backend. The setting is named `GGML_HEXAGON_VTCM_SIZE` and accepts K, M or G suffixes.
The release notes also show that a proposed `NTHREADS` cap was later dropped. They do not give a default VTCM size or explain how the change affects performance.
Why it matters
People running llama.cpp with Hexagon hardware can override the VTCM size, but the release notes do not say what value they should use or whether it improves performance.
