diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake
index 864f7fe320..1ee9d29932 100644
--- a/cmake/onnxruntime_python.cmake
+++ b/cmake/onnxruntime_python.cmake
@@ -608,6 +608,16 @@ if (onnxruntime_USE_CUDA)
     )
 endif()
 
+if (onnxruntime_USE_ROCM)
+    add_custom_command(
+      TARGET onnxruntime_pybind11_state POST_BUILD
+      COMMAND ${CMAKE_COMMAND} -E copy
+          $<TARGET_FILE:onnxruntime_providers_rocm>
+          $<TARGET_FILE:onnxruntime_providers_shared>
+          $<TARGET_FILE_DIR:${build_output_target}>/onnxruntime/capi/
+    )
+endif()
+
 if (onnxruntime_USE_TVM)
   add_custom_command(
     TARGET onnxruntime_pybind11_state POST_BUILD
diff --git a/setup.py b/setup.py
index 73c70f79b8..41b775b4f2 100644
--- a/setup.py
+++ b/setup.py
@@ -50,6 +50,7 @@ wheel_name_suffix = parse_arg_remove_string(sys.argv, '--wheel_name_suffix=')
 
 cuda_version = None
 rocm_version = None
+is_rocm = False
 # The following arguments are mutually exclusive
 if parse_arg_remove_boolean(sys.argv, '--use_tensorrt'):
     package_name = 'onnxruntime-gpu-tensorrt' if not nightly_build else 'ort-trt-nightly'
@@ -57,6 +58,7 @@ elif wheel_name_suffix == 'gpu':
     # TODO: how to support multiple CUDA versions?
     cuda_version = parse_arg_remove_string(sys.argv, '--cuda_version=')
 elif parse_arg_remove_boolean(sys.argv, '--use_rocm'):
+    is_rocm = True
     package_name = 'onnxruntime-rocm' if not nightly_build else 'ort-rocm-nightly'
     rocm_version = parse_arg_remove_string(sys.argv, '--rocm_version=')
 elif parse_arg_remove_boolean(sys.argv, '--use_openvino'):
@@ -155,7 +157,7 @@ try:
                 if len(args) > 3:
                     subprocess.run(args, check=True, stdout=subprocess.PIPE)
 
-                dest = 'onnxruntime/capi/libonnxruntime_providers_cuda.so'
+                dest = 'onnxruntime/capi/libonnxruntime_providers_' + 'rocm.so' if is_rocm else 'cuda.so'
                 if path.isfile(dest):
                     result = subprocess.run(['patchelf', '--print-needed', dest],
                                             check=True, stdout=subprocess.PIPE, universal_newlines=True)
@@ -190,17 +192,20 @@ except ImportError as error:
     print(error)
     bdist_wheel = None
 
+providers_cuda_or_rocm = 'libonnxruntime_providers_rocm.so' if is_rocm else 'libonnxruntime_providers_cuda.so'
+
 # Additional binaries
 if platform.system() == 'Linux':
     libs = ['onnxruntime_pybind11_state.so', 'libdnnl.so.2', 'libmklml_intel.so', 'libmklml_gnu.so', 'libiomp5.so',
             'mimalloc.so']
-    dl_libs = ['libonnxruntime_providers_shared.so', 'libonnxruntime_providers_cuda.so']
+    dl_libs = ['libonnxruntime_providers_shared.so']
+    dl_libs.append(providers_cuda_or_rocm)
     # DNNL, TensorRT & OpenVINO EPs are built as shared libs
     libs.extend(['libonnxruntime_providers_shared.so'])
     libs.extend(['libonnxruntime_providers_dnnl.so'])
     libs.extend(['libonnxruntime_providers_tensorrt.so'])
     libs.extend(['libonnxruntime_providers_openvino.so'])
-    libs.extend(['libonnxruntime_providers_cuda.so'])
+    libs.append(providers_cuda_or_rocm)
     # Nuphar Libs
     libs.extend(['libtvm.so.0.5.1'])
     if nightly_build: