Fix CI failure

Author: leonling-lly

include/cute/util/sycl_vec.hpp

@@ -59,10 +59,17 @@ using float2 = vector_t<float, 2>;
 using float4 = vector_t<float, 4>;
 using float8 = vector_t<float, 8>;
 
+#ifndef __SYCL_DEVICE_ONLY__
+using half = sycl::half;
+using half2 = vector_t<half, 2>;
+using half4 = vector_t<half, 4>;
+using half8 = vector_t<half, 8>;
+#else
 using half = _Float16;
 using half2 = vector_t<_Float16, 2>;
 using half4 = vector_t<_Float16, 4>;
 using half8 = vector_t<_Float16, 8>;
+#endif
 
 using short2 = vector_t<short, 2>;
 using short4 = vector_t<short, 4>;

tools/copy_debug/copy_debug.cpp

@@ -117,7 +117,7 @@ void copy(int global_M, int global_N) {
 
   auto gridDim = cutlasscompat::dim3(1);
   auto blockDim = cutlasscompat::dim3(SUBGROUP_SIZE);
-  launch<copy_kernel<CopyInstruction, decltype(tensor_S), fragment_size>, CopyKernelName<CopyInstruction, decltype(tensor_S), fragment_size>>(
+  launch<copy_kernel<CopyInstruction, decltype(tensor_S), fragment_size>, CopyKernelName<CopyInstruction, decltype(tensor_S)>>(
       launch_policy{gridDim, blockDim,
                     kernel_properties{sycl_exp::sub_group_size<SUBGROUP_SIZE>}},
       tensor_S);