meteor lake support not functionnal

i tried running python llamacpp on meteor lake core ultra 9 device on wsl with ubuntu 22.04 but observing some strange behavior

Expected behavior:
- build with sycl works correctly
- all devices are detected by llama cpp (Cpu, Npu and GPU as opencl, devices + GPU as sycl device)
- I am able to load gguf models (tinyllama Q4) on all devices
- run on cpu works well (around 58 tokens/s)

Odd behaviour:
- when loading the model on gpu, it suddently become very slow, with around 1.5 tokens/s

Any clue on what makes gpu run so slow?