diff --git a/build2cmake/src/templates/cuda/kernel.cmake b/build2cmake/src/templates/cuda/kernel.cmake
index 8205c959..c91e065c 100644
--- a/build2cmake/src/templates/cuda/kernel.cmake
+++ b/build2cmake/src/templates/cuda/kernel.cmake
@@ -49,7 +49,21 @@ if(GPU_LANG STREQUAL "CUDA")
   list(APPEND SRC {{'"${' + kernel_name + '_SRC}"'}})
 {% if supports_hipify %}
 elseif(GPU_LANG STREQUAL "HIP")
-  hip_archs_loose_intersection({{kernel_name}}_ARCHS "{{ rocm_archs|join(";") }}" ${ROCM_ARCHS})
+  hip_archs_loose_intersection({{kernel_name}}_ARCHS "{{ rocm_archs|join(";") }}" "${ROCM_ARCHS}")
+  message(STATUS "Archs for kernel {{kernel_name}}: {{ '${' + kernel_name + '_ARCHS}'}}")
+
+  foreach(_KERNEL_SRC {{'${' + kernel_name + '_SRC}'}})
+    if(_KERNEL_SRC MATCHES ".*\\.(cu|hip)$")
+      foreach(_ROCM_ARCH {{ '${' + kernel_name + '_ARCHS}'}})
+        set_property(
+          SOURCE ${_KERNEL_SRC}
+          APPEND PROPERTY
+          COMPILE_OPTIONS "$<$<COMPILE_LANGUAGE:HIP>:--offload-arch=${_ROCM_ARCH}>"
+        )
+      endforeach()
+    endif()
+  endforeach()
+
   list(APPEND SRC {{'"${' + kernel_name + '_SRC}"'}})
 {% endif %}
 endif()
diff --git a/build2cmake/src/templates/cuda/preamble.cmake b/build2cmake/src/templates/cuda/preamble.cmake
index 4abea735..78becb9a 100644
--- a/build2cmake/src/templates/cuda/preamble.cmake
+++ b/build2cmake/src/templates/cuda/preamble.cmake
@@ -11,7 +11,7 @@ message(STATUS "FetchContent base directory: ${FETCHCONTENT_BASE_DIR}")
 
 set(CUDA_SUPPORTED_ARCHS "{{ cuda_supported_archs }}")
 
-set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101")
+set(HIP_SUPPORTED_ARCHS "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201")
 
 include(${CMAKE_CURRENT_LIST_DIR}/cmake/utils.cmake)
 
@@ -85,11 +85,9 @@ if(GPU_LANG STREQUAL "CUDA")
 
   add_compile_definitions(CUDA_KERNEL)
 elseif(GPU_LANG STREQUAL "HIP")
-  set(ROCM_ARCHS "${HIP_SUPPORTED_ARCHS}")
-  # TODO: remove this once we can set specific archs per source file set.
-  override_gpu_arches(GPU_ARCHES
-    ${GPU_LANG}
-    "${${GPU_LANG}_SUPPORTED_ARCHS}")
+  override_gpu_arches(GPU_ARCHES HIP ${HIP_SUPPORTED_ARCHS})
+  set(ROCM_ARCHS ${GPU_ARCHES})
+  message(STATUS "ROCM supported target architectures: ${ROCM_ARCHS}")
 
   add_compile_definitions(ROCM_KERNEL)
 else()
diff --git a/build2cmake/src/templates/utils.cmake b/build2cmake/src/templates/utils.cmake
index dc522755..d4d4cef2 100644
--- a/build2cmake/src/templates/utils.cmake
+++ b/build2cmake/src/templates/utils.cmake
@@ -74,6 +74,7 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
   set(HIP_SRCS)
   foreach (SRC ${SRCS})
     get_source_file_property(include_dirs "${SRC}" INCLUDE_DIRECTORIES)
+    get_source_file_property(compile_options "${SRC}" COMPILE_OPTIONS)
     string(REGEX REPLACE "\.cu$" "\.hip" SRC ${SRC})
     string(REGEX REPLACE "cuda" "hip" SRC ${SRC})
 
@@ -84,6 +85,12 @@ function (hipify_sources_target OUT_SRCS NAME ORIG_SRCS)
         PROPERTIES INCLUDE_DIRECTORIES "${include_dirs}")
     endif()
 
+    if(compile_options)
+      set_source_files_properties(
+        ${SRC}
+        PROPERTIES COMPILE_OPTIONS "${compile_options}")
+    endif()
+
     list(APPEND HIP_SRCS "${CMAKE_CURRENT_BINARY_DIR}/${SRC}")
   endforeach()
 
@@ -516,8 +523,13 @@ function (define_gpu_extension_target GPU_MOD_NAME)
   endif()
 
   if (GPU_ARCHITECTURES)
-    set_target_properties(${GPU_MOD_NAME} PROPERTIES
-      ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
+    if (GPU_LANGUAGE STREQUAL "HIP")
+      # Clear target architectures, we are passing arch flags per source file.
+      set_property(TARGET ${GPU_MOD_NAME} PROPERTY HIP_ARCHITECTURES off)
+    else()
+      set_target_properties(${GPU_MOD_NAME} PROPERTIES
+        ${GPU_LANGUAGE}_ARCHITECTURES "${GPU_ARCHITECTURES}")
+    endif()
   endif()
 
   set_property(TARGET ${GPU_MOD_NAME} PROPERTY CXX_STANDARD 17)