b9667

ggml-org/llama.cppb9667Jun 16, 2026by github-actions[bot]same day as b9665

What changed since b9665

low risk

b9667 adds Vulkan backend support for gated_delta_net with a 16-element state vector, on top of b9665's offline bench flag.

New features

  • Vulkan backend can now execute the gated_delta_net activation function, limited to state vector size S_v=16.

Improvements

  • Broadens GPU-accelerated coverage of newer model architectures on Vulkan devices.

Upgrade notes

No configuration or API changes are required; the change is confined to the Vulkan backend. Users running gated_delta_net models with S_v=16 on Vulkan should see acceleration, while other state sizes remain on the prior fallback path. Note that b9665's --offline bench flag is already present in this build.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Extends Vulkan backend support to include the gated_delta_net activation function with a specific state vector size of 16.

Key Highlights

  • Support gated_delta_net activation function in Vulkan
  • Specifically adds support for S_v=16 state vector size

Release Notes (excerpt)

<details open> vulkan: Support gated_delta_net with S_v=16 (#24581) </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