b10456

ggml-org/llama.cppb10456Aug 17, 2026by github-actions[bot]same day as b10455

What changed since b10455

low risk

SYCL quantized copy kernel launch geometry fixed, boosting q4_0→f32 throughput roughly 8x on Arc GPUs.

Improvements

  • Thread and block counts for SYCL quantized copy kernels now scale with quant size, avoiding under- and over-subscription.
  • q4_0 to f32 conversion throughput on Arc A70 rose from about 20 GB/s to about 158 GB/s.
  • Other quantization types see flat-to-modest performance gains from the same launch adjustment.

Bug fixes

  • Corrected incorrect thread/block count calculation in SYCL quantized cpy kernel launches.

Upgrade notes

This is a single-commit follow-up to b10455 with no API, config, or model-format changes, so a straight binary swap is sufficient. Only SYCL users benefit; other backends are unaffected. The SYCL optimizer-step support from b10455 remains in place.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

This release fixes thread and block count calculations in quantized copy kernel launches for the SYCL backend, significantly improving throughput for certain quantization types.

Key Highlights

  • Fixed thread/block count in SYCL quantized cpy kernel launches
  • Significant performance improvement for q4_0 quantization
  • Improved performance for other quantization types

New Features

  • Performance optimization for SYCL quantized copy operations

Release Notes (excerpt)

<details open> sycl: fix thread/block count in quantized cpy kernel launches (#27160) Adjusts the thread/block count to be proportional to the size of the quant, reducing under/over subscription. Largest perf improvement is the q4_0 -> f32 path, with, on a Arc 70, throughput goes from 20.21 GB/s to 158.19 GB/s The rest of the quants are flat in performance uplift. </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories