b10306

ggml-org/llama.cppb10306Aug 7, 2026by github-actions[bot]same day as b10305

What changed since b10305

low risk

SYCL backend gains a contiguous fast path for fused GLU ops, speeding up split-operand SWIGLU workloads.

New features

  • SYCL fused GLU/SiLU kernels now dispatch a flat, de-interleave-free path when operands are contiguous and packed.
  • Added SWIGLU performance test cases covering 17408 columns with 512 and 2048 tokens in f16 and f32, both fused and split.

Improvements

  • Split-operand SWIGLU measured roughly 14% faster in f16 and 4% faster in f32 on Arc Pro B70; fused path unchanged.
  • Fused-GLU kernels were consolidated into a shared launcher that takes the activation op as an argument, removing duplicated code.
  • Removed the now-unused SYCL_GELU_BLOCK_SIZE and SYCL_SILU_BLOCK_SIZE constants without altering launch geometry.

Upgrade notes

This is a drop-in update with no config or API changes for normal users. Only out-of-tree SYCL code that referenced SYCL_GELU_BLOCK_SIZE or SYCL_SILU_BLOCK_SIZE needs adjusting, since those macros were deleted. The speedup applies to ggml_glu_split with packed operands; fused [gate|up] tensors still take the strided path.

Analyzed Oct 30, 2026

Upgrading?

Release notes summary

A performance-focused release for SYCL backends, specifically optimizing the GLU (Gated Linear Unit) operations. It introduces a flat path for contiguous memory and refactors kernel launches for better performance on Arc Pro B70.

Key Highlights

  • Added contiguous fast path for fused GLU/SiLU operations in SYCL (#26354)
  • Consolidated fused-GLU kernels to reduce code duplication
  • Performance improvements: +14% f16 and +4% f32 for split operands on Arc Pro B70
  • Added performance test cases for SWIGLU

Release Notes (excerpt)

<details open> sycl: *glu flat path (#26354) * tests: add SWIGLU perf cases perf mode had no GLU coverage. Adds SWIGLU at 17408 columns, 512 and 2048 tokens, f16 and f32, with the operands both fused and split. * sycl: consolidate fused-GLU kernels They differed only in which op_* they called, so take the op as an argument and share a common launcher. Their block sizes were all 256, so launch geometry is unchanged;…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories