diff --git a/sycl/include/sycl/detail/vector_arith.hpp b/sycl/include/sycl/detail/vector_arith.hpp
index e85e5e77cf97b..380691c0b1906 100644
--- a/sycl/include/sycl/detail/vector_arith.hpp
+++ b/sycl/include/sycl/detail/vector_arith.hpp
@@ -10,9 +10,6 @@
 
 #include <sycl/aliases.hpp>
 #include <sycl/detail/generic_type_traits.hpp>
-#include <sycl/detail/type_traits.hpp>
-#include <sycl/detail/type_traits/vec_marray_traits.hpp>
-#include <sycl/ext/oneapi/bfloat16.hpp>
 
 #include <functional>
 
diff --git a/sycl/include/sycl/detail/vector_convert.hpp b/sycl/include/sycl/detail/vector_convert.hpp
index 544250aff82ac..905334807ef93 100644
--- a/sycl/include/sycl/detail/vector_convert.hpp
+++ b/sycl/include/sycl/detail/vector_convert.hpp
@@ -54,8 +54,7 @@
 
 #pragma once
 
-#include <sycl/detail/generic_type_traits.hpp> // for is_sigeninteger, is_s...
-#include <sycl/exception.hpp>                  // for errc
+#include <sycl/detail/generic_type_traits.hpp>
 
 #include <sycl/detail/memcpy.hpp>
 #include <sycl/ext/oneapi/bfloat16.hpp>
@@ -63,7 +62,8 @@
 #include <sycl/vector.hpp>
 
 #ifndef __SYCL_DEVICE_ONLY__
-#include <cfenv> // for fesetround, fegetround
+#include <cfenv>
+#include <sycl/exception.hpp>
 #endif
 
 #include <type_traits>
diff --git a/sycl/include/sycl/ext/oneapi/experimental/cuda/builtins.hpp b/sycl/include/sycl/ext/oneapi/experimental/cuda/builtins.hpp
index 3609c282a5319..e0b8fbb861e0a 100644
--- a/sycl/include/sycl/ext/oneapi/experimental/cuda/builtins.hpp
+++ b/sycl/include/sycl/ext/oneapi/experimental/cuda/builtins.hpp
@@ -367,7 +367,7 @@ ldg(const T *ptr) {
   } else if constexpr (std::is_same_v<T, sycl::vec<half, 2>>) {
     typedef __fp16 h2 ATTRIBUTE_EXT_VEC_TYPE(2);
     auto rv = __nvvm_ldg_h2(reinterpret_cast<const h2 *>(ptr));
-    sycl::vec<half, 2> ret;
+    T ret;
     ret.x() = rv[0];
     ret.y() = rv[1];
     return ret;
@@ -376,7 +376,7 @@ ldg(const T *ptr) {
     h2 rv_2 = __nvvm_ldg_h2(reinterpret_cast<const h2 *>(ptr));
     auto rv = __nvvm_ldg_h(reinterpret_cast<const __fp16 *>(
         std::next(reinterpret_cast<const h2 *>(ptr))));
-    sycl::vec<half, 3> ret;
+    T ret;
     ret.x() = rv_2[0];
     ret.y() = rv_2[1];
     ret.z() = rv;
@@ -385,7 +385,7 @@ ldg(const T *ptr) {
     typedef __fp16 h2 ATTRIBUTE_EXT_VEC_TYPE(2);
     auto rv1 = __nvvm_ldg_h2(reinterpret_cast<const h2 *>(ptr));
     auto rv2 = __nvvm_ldg_h2(std::next(reinterpret_cast<const h2 *>(ptr)));
-    sycl::vec<half, 4> ret;
+    T ret;
     ret.x() = rv1[0];
     ret.y() = rv1[1];
     ret.z() = rv2[0];
diff --git a/sycl/include/sycl/marray.hpp b/sycl/include/sycl/marray.hpp
index 7419fbdcb73e6..e4c193b37b635 100644
--- a/sycl/include/sycl/marray.hpp
+++ b/sycl/include/sycl/marray.hpp
@@ -10,20 +10,15 @@
 
 #include <sycl/aliases.hpp>
 #include <sycl/detail/common.hpp>
-#include <sycl/detail/is_device_copyable.hpp>
-#include <sycl/half_type.hpp>
-
-#include <array>
-#include <cstddef>
-#include <cstdint>
-#include <type_traits>
-#include <utility>
+#include <sycl/detail/fwd/half.hpp>
 
 namespace sycl {
 inline namespace _V1 {
 
 template <typename DataT, std::size_t N> class marray;
 
+template <typename T> struct is_device_copyable;
+
 namespace detail {
 
 // Helper trait for counting the aggregate number of arguments in a type list,
diff --git a/sycl/include/sycl/vector.hpp b/sycl/include/sycl/vector.hpp
index 1d21930290a83..cf923d1a9f821 100644
--- a/sycl/include/sycl/vector.hpp
+++ b/sycl/include/sycl/vector.hpp
@@ -31,30 +31,16 @@
 #error "SYCL device compiler is built without ext_vector_type support"
 #endif
 
-#include <sycl/access/access.hpp>             // for decorated, address_space
-#include <sycl/aliases.hpp>                   // for half, cl_char, cl_int
-#include <sycl/detail/common.hpp>             // for ArrayCreator
-#include <sycl/detail/defines_elementary.hpp> // for __SYCL2020_DEPRECATED
-#include <sycl/detail/fwd/accessor.hpp>
-#include <sycl/detail/generic_type_traits.hpp> // for is_sigeninteger, is_s...
-#include <sycl/detail/memcpy.hpp>              // for memcpy
 #include <sycl/detail/named_swizzles_mixin.hpp>
-#include <sycl/detail/type_traits.hpp> // for is_floating_point
 #include <sycl/detail/vector_arith.hpp>
-#include <sycl/half_type.hpp> // for StorageT, half, Vec16...
 
-#include <sycl/ext/oneapi/bfloat16.hpp> // bfloat16
+#include <sycl/detail/common.hpp>
+#include <sycl/detail/fwd/accessor.hpp>
+#include <sycl/detail/fwd/half.hpp>
+#include <sycl/detail/memcpy.hpp>
 
-#include <algorithm>   // for std::min
-#include <array>       // for array
-#include <cassert>     // for assert
-#include <cstddef>     // for size_t, NULL, byte
-#include <cstdint>     // for uint8_t, int16_t, int...
-#include <functional>  // for divides, multiplies
-#include <iterator>    // for pair
-#include <ostream>     // for operator<<, basic_ost...
-#include <type_traits> // for enable_if_t, is_same
-#include <utility>     // for index_sequence, make_...
+#include <algorithm>
+#include <functional>
 
 namespace sycl {
 
@@ -63,6 +49,9 @@ namespace sycl {
 enum class rounding_mode { automatic = 0, rte = 1, rtz = 2, rtp = 3, rtn = 4 };
 
 inline namespace _V1 {
+namespace ext::oneapi {
+class bfloat16;
+}
 
 struct elem {
   static constexpr int x = 0;
@@ -512,8 +501,7 @@ class __SYCL_EBO vec :
 #endif
       bool, /*->*/ std::uint8_t,                            //
       sycl::half, /*->*/ sycl::detail::half_impl::StorageT, //
-      sycl::ext::oneapi::bfloat16,
-      /*->*/ sycl::ext::oneapi::bfloat16::Bfloat16StorageT, //
+      sycl::ext::oneapi::bfloat16, /*->*/ uint16_t,         //
       char, /*->*/ detail::ConvertToOpenCLType_t<char>,     //
       DataT, /*->*/ DataT                                   //
       >::type;