b10410

ggml-org/llama.cppb10410Aug 13, 2026by github-actions[bot]same day as b10408

What changed since b10408

low risk

SYCL gemm path is simplified to rely on automatic fp16 promotion, replacing the manual fp32 promotion in the non-oneDNN path.

Improvements

  • SYCL gemm now uses automatic fp16 promotion instead of explicit type promotion.
  • Removed the separate fp32 promotion branch from the non-oneDNN gemm path, reducing code duplication.
  • Minor internal cleanup of redundant comments in the SYCL backend.

Upgrade notes

This delta only touches the SYCL backend, so CPU, CUDA, Metal, and Vulkan users see no functional change. SYCL users should rebuild and sanity-check output quality/performance on Intel GPUs, since the gemm type-promotion path changed and could shift numerics slightly. No configuration or API changes are required.

Analyzed Oct 28, 2026

Upgrading?

Release notes summary

Refactors the SYCL gemm (General Matrix Multiply) implementation to use automatic fp16 promotion instead of manual type promotion.

Key Highlights

  • Remove separate fp32 type promotion in gemm non-oneDNN path (#26372)
  • Use automatic fp16 promotion in gemm

Release Notes (excerpt)

<details open> sycl: remove separate fp32 type promotion in gemm non-oneDNN path (#26372) * sycl: use automatic fp16 promotion in gemm * sycl: remove redundant comment </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)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories