b10427

ggml-org/llama.cppb10427Aug 14, 2026by github-actions[bot]same day as b10426

What changed since b10426

low risk

SYCL backend performance improved via dense FFN fusion optimization

New features

  • SYCL dense FFN fusion optimization for Q4_K models

Improvements

  • SYCL backend performance increased by 2.0% to 12.4% depending on batch size

Upgrade notes

No configuration changes are required. The SYCL backend optimization applies automatically to supported hardware. The WASI single-thread constraint from the previous release remains unchanged.

Analyzed Oct 16, 2026

Upgrading?

Release notes summary

The SYCL backend performance is further enhanced by fusing MulMat (gate) + MulMat (up) + GLU operations for Q4_K dense FFN.

Key Highlights

  • SYCL: Fuse mul_mat(gate) + mul_mat(up) + GLU for Q4_K dense FFN
  • Performance improvements ranging from +2.0% to +12.4% depending on batch size
  • Standard builds for all supported platforms

Release Notes (excerpt)

<details open> sycl: fuse mul_mat(gate) + mul_mat(up) + GLU for q4_K dense FFN (#26779) Measured on Arc Pro B70 (Battlemage, Level Zero), llama-bench -r 20, two interleaved rounds, tg128: qwen2.5-3B-Instruct Q4_K_M 154.18 -> 158.53 t/s +2.8% gemma-2-2b-it Q4_K_M 162.45 -> 165.62 t/s +2.0% llama-batched-bench on qwen2.5-3B, S_TG by batch size: B=1 142.72 -> 147.57 t/s +3.4% B=2 243.72 -> 268.26 t/s +10.1% B=4 359.58 -> 398.02 t/s +10.7%…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories