llama.cpp Build b10456 Landed This Morning With a SYCL Fix That Takes q4_0→f32 Throughput From 20.21 to 158.19 GB/s
GitHub / ggml-org/llama.cpp·high signal
Release b10456, published 2026-08-17 at 06:29 UTC, fixes the thread/block count in llama.cpp's quantized copy kernel launches on SYCL (PR #27160), which had been under- and over-subscribing the device depending on quant size. The measured result on an Arc B70 is a jump from 20.21 GB/s to 158.19 GB/s on the q4_0 → f32 path — roughly 7.8x — with the other quant types flat. If you run local inference on Intel GPUs through the SYCL backend, this is a one-build upgrade with a large dequantization win; CUDA and Metal users see nothing.