b10202

ggml-org/llama.cppb10202Jul 31, 2026by github-actions[bot]same day as b10201

What changed since b10201

low risk

SYCL backend now fuses RMS_NORM and MUL into one kernel, cutting kernel launches and memory traffic.

New features

  • SYCL backend gains a fused RMS_NORM + MUL kernel path for faster normalization-plus-scaling in transformer blocks.

Improvements

  • Fewer kernel launches and reduced intermediate memory traffic on SYCL devices due to the RMS norm/multiply fusion.

Upgrade notes

This is a backend-internal optimization with no CLI, API, or model-format changes, so a drop-in binary swap is sufficient. Only SYCL builds (Intel GPUs) benefit; other backends are unaffected. No rebuild flags or config changes are required.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

The SYCL backend has been optimized by fusing the RMS Normalization and Multiply operations into a single kernel.

Key Highlights

  • Fused RMS_NORM and MUL operations in the SYCL backend for improved performance

Release Notes (excerpt)

<details open> sycl: fuse RMS_NORM + MUL (#26015) </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 7.2) - 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