b10016

ggml-org/llama.cppb10016Jul 15, 2026by github-actions[bot]same day as b10015

What changed since b10015

medium risk

Adds SYCL Flash Attention via oneDNN XMX for Battlemage GPUs, boosting prefill speeds up to 4.26x at long context.

New features

  • SYCL Flash Attention path using the XMX engine through the oneDNN graph API, enabled with the existing flash-attention flag.
  • Multi-device synchronization support for the new SYCL FA path when more than one device is present.

Improvements

  • Prefill throughput on Battlemage (Xe2) improves roughly 1.21x at p=512 and 4.26x at p=80k for Qwen3.6-27b-Q8_0.
  • Reported ~32% gain in llama-bench pp512 with fa=1 and negligible perplexity change (~0.11%) on Llama-3.1-8B-Instruct Q8_0.
  • Non-Battlemage Intel architectures (e.g. Alchemist) automatically fall back to the existing FA kernel, preserving prior behavior.

Upgrade notes

No configuration or migration steps are required; the new path activates through the existing flash-attention option on supported SYCL hardware. Only Battlemage (Xe2) GPUs take the oneDNN XMX route, so other Intel devices see no behavioral change. Validate output quality if you rely on SYCL flash attention in multi-GPU setups, since the sync handling was newly added.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Introduces Flash Attention support for the SYCL backend using the XMX engine and oneDNN graph API, significantly improving prefill speeds on Battlemage architecture.

Key Highlights

  • SYCL Flash Attention with XMX engine via oneDNN graph API
  • Significant speedup: x1.21 at p=512 and x4.26 at p=80k for Qwen3.6-27b
  • Focused on Battlemage (BMG) architecture with multi-device sync support

Release Notes (excerpt)

<details open> [SYCL] Flash Attention with XMX engine via oneDNN (#25222) * [SYCL] F16 (default) Flash Attention with XMX engine via oneDNN graph API; Qwen3.6-27b-Q8_0 prefill speed up x1.21 at p=512 and x4.26 at p=80k * [SYCL] Address review on FA oneDNN path. Result: llama-bench---pp512; 32% increase with fa1; llama-perplexity---0.11% difference; tested model: mradermacher/Meta-Llama-3.1-8B-Instruct-Q8_0.gguf * PR-25222 revision v2: addressed audits…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories