Has anyone gotten Llama.cpp (or other) working using Intel iGPU (arrowlake) where it actually improves anything?

I Recently did a bunch of tests and wrote them all up on here, but the short version is that Vulkan basically doesn't work (or when it does, it's at 1tok/s at best). SYCL works pretty well, seems to run the Qwen3.6 35b models at around 12tok/s. The prefill part is around 20tok/s when it works but can be hit and miss and sometimes when you say "do that again" the llm just freezes up for ages. CPU by itself is the most consistent at actually working, and runs at more like 14tok/s on those models, and 30-40tok