b10069

ggml-org/llama.cppb10069Jul 20, 2026by github-actions[bot]2 days after b10068

What changed since b10068

low risk

b10069 adds OpenCL Adreno broadcast and view-offset support, improving llama-server multi-stream inference.

New features

  • Adreno OpenCL MUL_MAT kernels now support broadcast, enabling batched/multi-stream workloads.
  • Adreno Q8_0 MUL_MAT now honors view_offs, allowing correct operation on tensor views.
  • Broadcast support was extended to general OpenCL GEMM/GEMV paths, not just the no-shuffle variants.

Improvements

  • llama-server multi-stream handling on Adreno GPUs is more reliable thanks to the broadcast and view-offset fixes.
  • Removed redundant tests and stale comments from the OpenCL backend code.

Bug fixes

  • Fixes issue #25910 involving incorrect Adreno MUL_MAT results with broadcast inputs and non-zero view offsets.

Upgrade notes

No config or API changes are required; this is a drop-in update. Adreno GPU users running llama-server with multiple streams or K/V quantization should see correctness improvements. The earlier DFlash K/V cache rotation fix from b10068 remains included.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

OpenCL improvements for Adreno GPU support including broadcast and view offsets.

Key Highlights

  • Support broadcast for Adreno MUL_MAT operations
  • Honor view_offs for Adreno Q8_0 MUL_MAT
  • Improves llama-server multi-stream support
  • Fixes issue #25910

Release Notes (excerpt)

<details open> opencl: Support broadcast for Adreno MUL_MAT and honor `view_offs` for Adreno Q8_0 MUL_MAT for llama-server multi-stream (#25910) * opencl: handle broadcast for adreno gemm/gemv_noshuffle * opencl: honor view_offs for adreno noshuffle gemm/gemv * opencl: general GEMM/GEMV support broadcast * opencl: remove unnecessary tests * opencl: remove unnecessary comments --------- Co-authored-by: Li He <lih@qti.qualcomm.com> </details> **Website:** - <https://llama.app>…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories