b10411

ggml-org/llama.cppb10411Aug 13, 2026by github-actions[bot]same day as b10410

What changed since b10410

low risk

b10411 adds a vectorized F16→F32 conversion for the flash-attention V-cache in the CPU backend; no other changes.

Improvements

  • Flash-attention on the CPU backend now converts V-cache entries from F16 to F32 using vectorized code instead of scalar loops.
  • Faster prompt processing and token generation for F16 KV-cache flash-attention workloads on CPU.

Upgrade notes

This is a single-commit performance patch on top of b10410, with no API, config, or model-format changes. Users running CPU flash-attention with an F16 V-cache should see a modest speedup; everyone else is unaffected. The SYCL gemm fp16 promotion refactor landed in b10410, not here, so no additional action is needed for SYCL builds.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

Optimizes the Flash-Attention implementation by vectorizing the V-cache F16 to F32 conversion process.

Key Highlights

  • Vectorize flash-attention V-cache F16 to F32 conversion (#26947)

Release Notes (excerpt)

<details open> ggml-cpu/ops: vectorize flash-attention V-cache F16 to F32 conversion (#26947) Co-authored-by: jinzihao <jinzihao.jzh@alibaba-inc.com> </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) - Ubuntu arm64 (CPU) - Ubuntu s390x (CPU) - Ubuntu x64 (Vulkan) - Ubuntu arm64 (Vulkan) - Ubuntu x64 (ROCm 7.14)DISABLED…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories