b9933

ggml-org/llama.cppb9933Jul 8, 2026by github-actions[bot]same day as b9932

What changed since b9932

low risk

b9933 is a targeted OpenCL Q6_K correctness fix over b9932's Vulkan tuning, with no API changes.

Improvements

  • Reserves alignment slack when carving SOA subbuffers to avoid overlapping adjacent tensors.
  • Routes misaligned Q6_K dense matmul shapes away from the noshuffle path on OpenCL.
  • Keeps standard 128-multiple dimensions on the existing noshuffle fast path.

Bug fixes

  • Fixes garbled OpenCL output for Q6_K weights where ne01 is not a multiple of 128.
  • Prevents silent corruption when misaligned quant component subbuffers overlap neighboring tensors.

Upgrade notes

No configuration or migration changes are required. Rebuild or update binaries if you use OpenCL with Q6_K models whose dimensions are not multiples of 128, especially on Adreno.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Fixes garbled output in OpenCL for Q6_K weights with specific alignment requirements.

Key Highlights

  • Fixed garbled output for Q6_K weights where ne01 is not a multiple of 128
  • Added alignment slack reservation for SOA subbuffer carving
  • Implemented fallback path using LM-based Q6_K mm for misaligned shapes

New Features

  • Fix for Q6_K GEMM/GEMV alignment issues on OpenCL

Release Notes (excerpt)

<details open> opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (#25464) * opencl: fix garbled output for Q6_K weights with ne01 % 128 != 0 on Adreno Observed with granite-3.1-3b-a800m-instruct, whose vocab is an odd number. Route Q6_K dense mul_mat with ne01 % 128 != 0 off the noshuffle path: decode (ne1==1) uses the correct flat GEMV and the matching GEMM (ne1>1) falls back to CPU (the flat convert has no verified small-batch GEMM kernel…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories