b9668

ggml-org/llama.cppb9668Jun 16, 2026by github-actions[bot]same day as b9667

What changed since b9667

low risk

b9668 adds a Vulkan optimization preferring host-visible memory buffers on UMA devices, on top of b9667's gated_delta_net S_v=16 support.

New features

  • Vulkan backend now implements UMA host-visible memory handling for unified memory architectures.

Improvements

  • Vulkan allocator prefers host-visible memory buffers on UMA devices, reducing memory management overhead.
  • Better memory efficiency and performance on integrated/shared-memory GPUs such as those in laptops and handhelds.

Upgrade notes

No configuration or API changes are required; the UMA behavior is applied automatically when the Vulkan backend detects a unified memory device. Users on discrete GPUs should see no behavioral difference, while UMA users may notice changed memory allocation patterns worth benchmarking.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

Optimizes Vulkan performance on Unified Memory Architecture (UMA) devices by implementing a preference for host-visible memory buffers to improve memory management efficiency.

Key Highlights

  • Implement UMA host-visible memory
  • Prefer host-visible memory buffers on UMA devices

Release Notes (excerpt)

<details open> vulkan: prefer host-visible memory buffers on UMA devices (#22930) * implement UMA host-visible memory * update based on 0cc4m's suggestion </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories