b9946

ggml-org/llama.cppb9946Jul 9, 2026by github-actions[bot]same day as b9945

What changed since b9945

low risk

Hexagon backend gains tiling, fast division, and kernel optimizations; no changes outside Qualcomm builds.

New features

  • Pointwise unary ops on the Hexagon backend now tile wide rows to avoid VTCM overflow.
  • Hexagon unary kernels use fastdiv instead of plain division for faster execution.
  • Added VTCM layout handling with host-computed kernel parameters for Hexagon unary ops.
  • Introduced tracing instrumentation for Hexagon unary operations.
  • Specialized thread functions were added to improve generated Hexagon code.

Improvements

  • Hexagon HVX kernels were factored out and de-duplicated for cleaner, more maintainable code.
  • Fastdiv initialization was moved into kernel parameters to reduce per-call overhead.
  • CMake configuration for Hexagon was adjusted to support LTO builds.
  • Permuted tensors are now explicitly rejected in unary ops since models don't use them.

Bug fixes

  • Fixed a std::min collision with the Windows min macro in the Hexagon backend.
  • Prevented VTCM overflow on wide-row pointwise unary operations via tiling.

Upgrade notes

This release only affects the Hexagon (Qualcomm) backend; CPU, CUDA, Metal, and other backends are unchanged. No configuration or API changes are required, so upgrading is a drop-in replacement. Users not targeting Hexagon hardware will see no behavioral difference.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Added optimizations for Hexagon (Qualcomm) backend including tiling and fast division.

Key Highlights

  • Added tiling for wide rows in pointwise unary ops to avoid VTCM overflow
  • Replaced division operations with fastdiv in hex-unary
  • Added vtcm layout and host-computed kernel parameters
  • Added specialized thread functions to improve generated code
  • Fixed std::min collision with Windows min macro

Release Notes (excerpt)

<details open> hexagon: tiling, tracing and optimizations for unary ops (#25474) * hexagon: tile wide rows in pointwise unary ops to avoid VTCM overflow * unary: reject permuted tensors for now (not used by models) * hex-unary: replace divs with fastdiv * hex-unary: add vtcm layout and host computed kernel params * hex-unary: move fastdiv init into kernel params * hex-unary: add specialized thread functions to improve generated code * hex-unary: tracing instrumentation for unary ops…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories