The ggml-org team has released llama.cpp b11849, adding support for the SGN operation on Hexagon, Qualcomm’s processor architecture. The release notes say the change includes half-precision support and an optimization that uses the sign bit.
The release also provides builds for macOS, iOS and Ubuntu, including CPU builds for x64, arm64 and s390x. The notes do not say which devices or workloads will see a practical speed or performance change.
Why it matters
Developers using llama.cpp with Hexagon hardware can use the new SGN operation, but the release notes do not quantify its effect on model performance.
