b9717

ggml-org/llama.cppb9717Jun 19, 2026by github-actions[bot]same day as b9716

What changed since b9716

low risk

b9717 adds Power10 CPU MMA support for K-dimension tails in Q8/Q4 matmul, reducing fallback to mnpack.

New features

  • Power10 tinyBlas Q8/Q4 MMA matmul now handles K dimensions not divisible by kc by processing the final K panel at its actual depth.

Improvements

  • More workloads can now use the Power10 MMA kernel instead of falling back to the slower mnpack path.
  • Packing and kernel execution accept a reduced final panel size, improving matmul efficiency on Power10 CPUs.

Upgrade notes

This is a Power10-specific ggml-cpu optimization with no API or configuration changes; existing binaries and model files remain compatible. Users on non-Power10 hardware will see no behavioral difference. Simply rebuild or pull the newer release to benefit.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Optimizes Power10 CPU matmul performance by removing the requirement for the K dimension to be divisible by kc in Q8/Q4 MMA kernels.

Key Highlights

  • Supports K-tails in Power10 Q8/Q4 matmul
  • Reduces requirement for K divisibility
  • Reduces fallback to mnpack

Release Notes (excerpt)

<details open> ggml-cpu: support K tails in power10 Q8/Q4 MMA matmul (#24753) * ggml-cpu: support K tails in Power10 MMA Q8/Q4 matmul This patch removes the requirement that K be divisible by kc in the tinyBlas_Q0_PPC tiled matmul path. Process the final K panel using its actual depth and pass the reduced panel size through packing and kernel execution. This allows more workloads to use the MMA kernel and reduces fallback to mnpack. * Apply suggestion from @taronaeo…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories