llama.cpp Adds Intel SYCL Support and Performance Improvements

·2026.07.15 17:52

Key point

llama.cpp has rolled out an update adding Flash Attention support on Intel SYCL and improving Qwen model inference speed.

llama.cpp's SYCL/Intel support has been significantly enhanced. The key updates are as follows.

Flash Attention Support: Through the oneDNN graph API (SDPA), XMX engine-based Flash Attention is now supported on the Xe2 architecture.
Qwen Model Performance Improvement: The prefill speed of the Qwen3.6-27b-Q8_0 model has improved by 1.21x (p=512) up to 4.26x (p=80k) depending on sequence length.
Other Optimizations: These include an increased minimum buffer size for USM system allocation, OP XIELU support, fp16 support for the conv2d_dw kernel, and bug fixes for certain quantization methods (Q2_K, Q4_K, Q5_K).

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.