diff --git a/cmake/onnxruntime_python.cmake b/cmake/onnxruntime_python.cmake
index 062cc8f9dbff3..07c65e7986b05 100644
--- a/cmake/onnxruntime_python.cmake
+++ b/cmake/onnxruntime_python.cmake
@@ -667,6 +667,15 @@ add_custom_command(
       $<TARGET_FILE_DIR:${build_output_target}>
 )
 
+if (onnxruntime_BUILD_SHARED_LIB)
+  add_custom_command(
+    TARGET onnxruntime_pybind11_state POST_BUILD
+    COMMAND ${CMAKE_COMMAND} -E copy
+        $<TARGET_FILE:onnxruntime>
+        $<TARGET_FILE_DIR:${build_output_target}>/onnxruntime/capi/
+  )
+endif()
+
 if (onnxruntime_USE_OPENVINO)
   add_custom_command(
     TARGET onnxruntime_pybind11_state POST_BUILD
diff --git a/setup.py b/setup.py
index 3203993e0c4d4..5750833ce35de 100644
--- a/setup.py
+++ b/setup.py
@@ -297,11 +297,13 @@ def finalize_options(self):
         "libmklml_gnu.so",
         "libiomp5.so",
         "mimalloc.so",
+        "libonnxruntime.so*",
     ]
     dl_libs = ["libonnxruntime_providers_shared.so"]
     dl_libs.append(providers_cuda_or_rocm)
     dl_libs.append(providers_tensorrt_or_migraphx)
     dl_libs.append(providers_cann)
+    dl_libs.append("libonnxruntime.so*")
     # DNNL, TensorRT & OpenVINO EPs are built as shared libs
     libs.extend(["libonnxruntime_providers_shared.so"])
     libs.extend(["libonnxruntime_providers_dnnl.so"])
@@ -313,7 +315,12 @@ def finalize_options(self):
     if nightly_build:
         libs.extend(["libonnxruntime_pywrapper.so"])
 elif platform.system() == "Darwin":
-    libs = ["onnxruntime_pybind11_state.so", "libdnnl.2.dylib", "mimalloc.so"]  # TODO add libmklml and libiomp5 later.
+    libs = [
+        "onnxruntime_pybind11_state.so",
+        "libdnnl.2.dylib",
+        "mimalloc.so",
+        "libonnxruntime.dylib*",
+    ]  # TODO add libmklml and libiomp5 later.
     # DNNL & TensorRT EPs are built as shared libs
     libs.extend(["libonnxruntime_providers_shared.dylib"])
     libs.extend(["libonnxruntime_providers_dnnl.dylib"])
@@ -323,7 +330,13 @@ def finalize_options(self):
     if nightly_build:
         libs.extend(["libonnxruntime_pywrapper.dylib"])
 else:
-    libs = ["onnxruntime_pybind11_state.pyd", "dnnl.dll", "mklml.dll", "libiomp5md.dll"]
+    libs = [
+        "onnxruntime_pybind11_state.pyd",
+        "dnnl.dll",
+        "mklml.dll",
+        "libiomp5md.dll",
+        "onnxruntime.dll",
+    ]
     # DNNL, TensorRT & OpenVINO EPs are built as shared libs
     libs.extend(["onnxruntime_providers_shared.dll"])
     libs.extend(["onnxruntime_providers_dnnl.dll"])
@@ -376,7 +389,7 @@ def finalize_options(self):
         dl_libs.append("plugins.xml")
         dl_libs.append("usb-ma2x8x.mvcmd")
     data = ["capi/libonnxruntime_pywrapper.so"] if nightly_build else []
-    data += [path.join("capi", x) for x in dl_libs if path.isfile(path.join("onnxruntime", "capi", x))]
+    data += [path.join("capi", x) for x in dl_libs if glob(path.join("onnxruntime", "capi", x))]
     ext_modules = [
         Extension(
             "onnxruntime.capi.onnxruntime_pybind11_state",
@@ -384,7 +397,7 @@ def finalize_options(self):
         ),
     ]
 else:
-    data = [path.join("capi", x) for x in libs if path.isfile(path.join("onnxruntime", "capi", x))]
+    data = [path.join("capi", x) for x in libs if glob(path.join("onnxruntime", "capi", x))]
     ext_modules = []
 
 # Additional examples
diff --git a/tools/ci_build/build.py b/tools/ci_build/build.py
index b73a17db3ce13..ae4c9b27544ba 100644
--- a/tools/ci_build/build.py
+++ b/tools/ci_build/build.py
@@ -2593,7 +2593,16 @@ def main():
     if args.build_wheel or args.gen_doc or args.use_tvm or args.enable_training:
         args.enable_pybind = True
 
-    if args.build_csharp or args.build_nuget or args.build_java or args.build_nodejs:
+    if (
+        args.build_csharp
+        or args.build_nuget
+        or args.build_java
+        or args.build_nodejs
+        or (args.enable_pybind and not args.enable_training)
+    ):
+        # If pyhon bindings are enabled, we embed the shared lib in the python package.
+        # If training is enabled, we don't embed the shared lib in the python package since training requires
+        # torch interop.
         args.build_shared_lib = True
 
     if args.build_nuget and cross_compiling: