b9934

ggml-org/llama.cppb9934Jul 9, 2026by github-actions[bot]same day as b9933

What changed since b9933

low risk

b9934 adds WebGPU flash-attention subgroup-split tuning on top of b9933's OpenCL Q6_K fix, with no API changes.

New features

  • WebGPU flash_attn_vec now tunes the subgroup split (d_split) parameter for improved vector operation efficiency.

Improvements

  • WebGPU Flash Attention vector kernels are better tuned via d_split, potentially improving throughput on WebGPU backends.
  • The OpenCL Q6_K alignment fix from b9933 (misaligned ne01 handling and SOA subbuffer slack) remains in place.

Upgrade notes

This is a routine incremental build: the only delta is a WebGPU kernel tuning change, and the OpenCL Q6_K correctness fix from b9933 carries over. No configuration, model, or API changes are required; users on OpenCL with non-128-multiple vocab dims should ensure they are at least on b9933 or later. WebGPU users may see modest Flash Attention performance differences.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

Tuning of subgroup splitting in WebGPU for better Flash Attention vector operations.

Key Highlights

  • Tuned subgroup split (d_split) in flash_attn_vec

Release Notes (excerpt)

<details open> ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec (#25418) </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories