Currently, scattered tensors in slm are not supported for load/store ops:
#xegpu.tdesc_attr<scattered = true, memory_scope=slm>>
They can only load/store to global memory (ugm), as indicated by the hardcoded sfid argument of the intrinsic call which is set to 15. Changing it to 14 should allow loads/stores to be performed upon shared local memory (slm).