b10201

ggml-org/llama.cppb10201Jul 31, 2026by github-actions[bot]1 days after b10199

What changed since b10199

low risk

WebGPU Flash Attention vector path optimized for quantized KV caches at long context lengths.

Improvements

  • WebGPU flash_attn_vec now handles quantized KV cache more efficiently when context lengths grow large.
  • Cleanup of type checks and comments in the WebGPU Flash Attention code, plus a rebase-induced build error resolved.

Bug fixes

  • Corrected several bugs and a vector-type check in the WebGPU quantized-KV Flash Attention path.

Upgrade notes

This is a build-level change confined to the WebGPU backend; no configuration or API changes are required. Users on other backends (CPU, CUDA, Vulkan, ROCm) will see no behavioral difference and can upgrade freely. Rebuild or pull the new WebGPU binaries to pick up the long-context speedup.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

This release optimizes the WebGPU implementation of Flash Attention for quantized Key-Value (KV) caches, specifically improving performance at long context lengths.

Key Highlights

  • Improved Flash Attention vector operations for quantized KV cache at long contexts
  • Fixed bugs and type checks in the implementation

New Features

  • Optimized WebGPU Flash Attention for long contexts

Release Notes (excerpt)

<details open> ggml-webgpu: improve flash_attn_vec for quantized KV at long contexts (#25956) * improve fa of quantized kv cache * Fix some bugs and some comments. * fix v type check and some comments * Fix build error caused by rebasing * editorconfig checking pass </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:** - Ubuntu x64 (CPU)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories