b10052

ggml-org/llama.cppb10052Jul 16, 2026by github-actions[bot]same day as b10051

What changed since b10051

medium risk

Hexagon backend overhauled with dirty-bit L2 cache tracking, lazy/threaded flushes, and workqueue API rework.

New features

  • Hexagon L2 cache now tracks dirty/clean tensor state and delays or skips flushes based on alias overlap.
  • Dedicated main threads plus explicit wakeup/suspend with spinning threads in the Hexagon work-queue.
  • DMA queue support for alias queues and cached DMA transfers.
  • Tiled activation processing for HMX matmul to better distribute HVX work.
  • Tracing instrumentation for L2 flush events and matmul initialization.

Improvements

  • L2 flushes are optimized at the opbatch level instead of per-operation.
  • Legacy workpool API reworked to align with the hmx-queue and dma-queue interfaces.
  • Tensor alias handling simplified, with overlapping views modeled as a circular alias list.
  • Non-weight tensors are now consistently marked as compute buffers in the Hexagon backend.
  • General cleanup of work-queue APIs and removal of unused masks and redundant HVX context drops.

Bug fixes

  • Removed an artificial solver limit that capped the number of activation-prep threads.
  • Fixed f32 activation buffer alignment for nhvx values 5, 6, and 7.
  • Fixed a compiler warning in the Hexagon matmul path.
  • The --top option no longer incorrectly applies to the hex-prof timeline report.

Upgrade notes

This delta is confined to the Hexagon backend; no CLI flags, model formats, or config files change, so a rebuild of the binaries is sufficient. Downstream patches that touch the legacy workpool API will need to be ported to the new hmx-queue/dma-queue-style interface. Non-Hexagon platforms should see no behavioral difference from b10051.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Major overhaul of Hexagon L2 cache handling, introducing dirty bit tracking, threaded flushing, and various MUL_MAT and workqueue improvements.

Key Highlights

  • Reworked L2 cache handling with lazy flushing and dirty bit tracking.
  • Introduced threaded L2 flushes and dedicated main threads.
  • Improved tensor alias handling, DMA queue support, and workqueue APIs.

Release Notes (excerpt)

<details open> hexagon: L2 cache handling rework (dirty bit tracking with lazy flushing) and more MUL_MAT updates (#25762) * hex-mm: fix artificial limit in the solver that restricted number of act-prep threads * hex-mm: fix warning * hex-prof: do not apply --top to the timeline report * hmx-mm: add suport for tiled act-processing to better distribute hvx work * hex-l2: add tracing for l2flush events * workqueue: redo the legacy workpool api to match hmx-queue and dma-queue…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories