b9678

ggml-org/llama.cppb9678Jun 17, 2026by github-actions[bot]same day as b9677

What changed since b9677

low risk

Single OpenCL kernel optimization for decode-phase matmul; no functional or API changes between b9677 and b9678.

Improvements

  • OpenCL mul_mat_f16_f32_l4 kernel tuned for the decode phase, improving token-generation throughput on OpenCL devices.
  • Performance-only change scoped to the OpenCL backend; CPU, CUDA, Metal, Vulkan, and ROCm paths are unaffected.

Upgrade notes

Drop-in replacement with no config, model, or API changes; only users running the OpenCL backend will see any difference. Rebuild or re-download the OpenCL-enabled binaries to pick up the faster kernel. No migration steps required.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

This release optimizes the OpenCL backend's matrix multiplication operation (`mul_mat_f16_f32_l4`) specifically for the decode phase, likely improving performance for inference tasks.

Key Highlights

  • Optimization of OpenCL `mul_mat_f16_f32_l4` for decode phase
  • Performance improvements for inference on OpenCL devices

New Features

  • OpenCL decode phase optimization

Release Notes (excerpt)

<details open> opencl: optimize mul_mat_f16_f32_l4 for decode (#24504) </details> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.2) - Ubuntu x64 (OpenVINO) - Ubuntu x64 (SYCL FP32) - Ubuntu x64 (SYCL FP16) **Android:** - Android arm64 (CPU) **Windows:**…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories