b9650

ggml-org/llama.cppb9650Jun 15, 2026by github-actions[bot]same day as b9649

What changed since b9649

low risk

Single SYCL fix corrects the max-reduction logic in soft_max_f32, with no other changes between b9649 and b9650.

Bug fixes

  • Fixed incorrect max-reduction logic in the SYCL soft_max_f32 kernel, restoring correct softmax math on SYCL devices.

Upgrade notes

This is a one-commit patch on top of b9649; only SYCL users are affected, and no config or API changes are needed. Rebuild or pull the updated SYCL binaries if you rely on softmax-heavy workloads on Intel GPUs.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

This release corrects a bug in the SYCL soft_max_f32 function where the max reduction logic was incorrect.

Key Highlights

  • SYCL: Fixed soft_max_f32 max reduction
  • Corrected mathematical logic for softmax operations

New Features

  • SYCL soft_max_f32 max reduction fix

Release Notes (excerpt)

<details open> sycl: fix soft_max_f32 max reduction (#24451) </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