b9885

ggml-org/llama.cppb9885Jul 6, 2026by github-actions[bot]same day as b9884

What changed since b9884

low risk

b9885 enables tiled matrix multiplication on AIX with smaller stack buffers, fixing segfaults and roughly doubling prompt-processing speed there.

New features

  • Tiled matmul path is now usable on AIX systems, where it was previously disabled due to crashes.

Improvements

  • Reduced local stack buffer footprint for A_pack and B_pack in the tiled matmul path to avoid stack exhaustion on AIX.
  • Reported ~2x PP_Speed gains on AIX for FP32, Q4_0 and Q8_0 models in llama-bench, llama-batched-bench and llama-cli.

Bug fixes

  • Segmentation faults on AIX caused by oversized stack buffers in matmul_tiled are resolved.

Upgrade notes

This is a platform-specific patch; non-AIX users should see no behavioral change and can upgrade without config or model changes. AIX operators get the tiled path back automatically, so benchmark prompt-processing throughput after upgrading to confirm the expected speedup.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Tiled matrix multiplication has been enabled for AIX systems by reducing local stack buffer footprints to avoid segmentation faults. This change delivers approximately 2x performance improvements for specific model types on the platform.

Key Highlights

  • Enable tiled matmul on AIX
  • Fix segmentation fault on AIX
  • Performance boost of ~2x for PP_Speed

Release Notes (excerpt)

<details open> ggml-cpu: Enable tiled matmul on AIX (#25199) The matmul_tiled path uses large local stack buffers for A_pack and B_pack. On AIX this can trigger a segmentation fault, so reduce the buffer footprint there to keep the tiled path usable. Performance Impact: ~ 2x gains in PP_Speed for FP32, Q4_0 and Q8_0 models tested with llama-bench, llama-batched-bench and llama-cli. Models used: Llama3.2 3b Instruct F32, qwen 2.5 3b Q4_0 and Q8_0 </details> **macOS/iOS:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories