The ggml-org/llama.cpp b11979 release changes how its SYCL kernels convert several quantized model formats on Intel GPUs. On an Arc Pro B70, the project measured faster prefill—the processing of prompt text—when tensors used a reordered layout, with gains varying by format and model.
The largest reported gain on that path was 28.7% for a Q6_K 27-billion-parameter model. For MXFP4 mixture-of-experts models, prefill improved 57% to 61% in the reported test, including with a fresh layout; decode speed was unchanged. The measurements are project-reported, and perplexity results were unchanged or within run-to-run noise.
Why it matters
People running these model formats on an Intel Arc Pro B70 may see shorter prompt-processing time, though the gains depend on the model and layout and do not speed up decoding.
