diff --git a/tools/dynamo/verify_dynamo.py b/tools/dynamo/verify_dynamo.py
index 364773b7e1884d..ba2e59e557e224 100644
--- a/tools/dynamo/verify_dynamo.py
+++ b/tools/dynamo/verify_dynamo.py
@@ -5,10 +5,8 @@
 import traceback
 import warnings
 
-from pkg_resources import packaging
-
-MIN_CUDA_VERSION = packaging.version.parse("11.6")
-MIN_ROCM_VERSION = packaging.version.parse("5.4")
+MIN_CUDA_VERSION = "11.6"
+MIN_ROCM_VERSION = "5.4"
 MIN_PYTHON_VERSION = (3, 8)
 
 
@@ -28,11 +26,12 @@ def check_python():
 def check_torch():
     import torch
 
-    return packaging.version.parse(torch.__version__)
+    return torch.__version__
 
 
 # based on torch/utils/cpp_extension.py
 def get_cuda_version():
+    from torch.torch_version import TorchVersion
     from torch.utils import cpp_extension
 
     CUDA_HOME = cpp_extension._find_cuda_home()
@@ -50,10 +49,11 @@ def get_cuda_version():
         raise VerifyDynamoError("CUDA version not found in `nvcc --version` output")
 
     cuda_str_version = cuda_version.group(1)
-    return packaging.version.parse(cuda_str_version)
+    return TorchVersion(cuda_str_version)
 
 
 def get_rocm_version():
+    from torch.torch_version import TorchVersion
     from torch.utils import cpp_extension
 
     ROCM_HOME = cpp_extension._find_rocm_home()
@@ -75,16 +75,17 @@ def get_rocm_version():
 
     hip_str_version = hip_version.group(1)
 
-    return packaging.version.parse(hip_str_version)
+    return TorchVersion(hip_str_version)
 
 
 def check_cuda():
     import torch
+    from torch.torch_version import TorchVersion
 
     if not torch.cuda.is_available() or torch.version.hip is not None:
         return None
 
-    torch_cuda_ver = packaging.version.parse(torch.version.cuda)
+    torch_cuda_ver = TorchVersion(torch.version.cuda)
 
     # check if torch cuda version matches system cuda version
     cuda_ver = get_cuda_version()
@@ -112,14 +113,13 @@ def check_cuda():
 
 def check_rocm():
     import torch
+    from torch.torch_version import TorchVersion
 
     if not torch.cuda.is_available() or torch.version.hip is None:
         return None
 
     # Extracts main ROCm version from full string
-    torch_rocm_ver = packaging.version.parse(
-        ".".join(list(torch.version.hip.split(".")[0:2]))
-    )
+    torch_rocm_ver = TorchVersion(".".join(list(torch.version.hip.split(".")[0:2])))
 
     # check if torch rocm version matches system rocm version
     rocm_ver = get_rocm_version()
diff --git a/torch/utils/cpp_extension.py b/torch/utils/cpp_extension.py
index 34bebe6af5fe8e..9cf602206bf8e2 100644
--- a/torch/utils/cpp_extension.py
+++ b/torch/utils/cpp_extension.py
@@ -25,7 +25,6 @@
 from torch.torch_version import TorchVersion
 
 from setuptools.command.build_ext import build_ext
-from pkg_resources import packaging  # type: ignore[attr-defined]
 
 IS_WINDOWS = sys.platform == 'win32'
 IS_MACOS = sys.platform.startswith('darwin')
@@ -402,9 +401,11 @@ def _check_cuda_version(compiler_name: str, compiler_version: TorchVersion) -> N
     if cuda_version is None:
         return
 
-    cuda_str_version = cuda_version.group(1)
-    cuda_ver = packaging.version.parse(cuda_str_version)
-    torch_cuda_version = packaging.version.parse(torch.version.cuda)
+    cuda_ver = cuda_str_version = cuda_version.group(1)
+    if torch.version.cuda is None:
+        return
+
+    torch_cuda_version = TorchVersion(torch.version.cuda)
     if cuda_ver != torch_cuda_version:
         # major/minor attributes are only available in setuptools>=49.4.0
         if getattr(cuda_ver, "major", None) is None:
@@ -1133,7 +1134,7 @@ def CUDAExtension(name, sources, *args, **kwargs):
         extra_compile_args_dlink += [f'-L{x}' for x in library_dirs]
         extra_compile_args_dlink += [f'-l{x}' for x in dlink_libraries]
 
-        if (torch.version.cuda is not None) and packaging.version.parse(torch.version.cuda) >= packaging.version.parse('11.2'):
+        if (torch.version.cuda is not None) and TorchVersion(torch.version.cuda) >= '11.2':
             extra_compile_args_dlink += ['-dlto']   # Device Link Time Optimization started from cuda 11.2
 
         extra_compile_args['nvcc_dlink'] = extra_compile_args_dlink
@@ -1449,7 +1450,9 @@ def build_precompile_header(pch_cmd):
             raise RuntimeError(f"Compile PreCompile Header fail, command: {pch_cmd}") from e
 
     extra_cflags_str = listToString(extra_cflags)
-    extra_include_paths_str = " ".join([f'-I{include}' for include in extra_include_paths])
+    extra_include_paths_str = (
+        "" if extra_include_paths is None else " ".join([f"-I{include}" for include in extra_include_paths])
+    )
 
     lib_include = os.path.join(_TORCH_PATH, 'include')
     torch_include_dirs = [
@@ -2344,9 +2347,8 @@ def sanitize_flags(flags):
         # --generate-dependencies-with-compile was added in CUDA 10.2.
         # Compilation will work on earlier CUDA versions but header file
         # dependencies are not correctly computed.
-        required_cuda_version = packaging.version.parse('11.0')
-        has_cuda_version = torch.version.cuda is not None
-        if has_cuda_version and packaging.version.parse(torch.version.cuda) >= required_cuda_version:
+        required_cuda_version = '11.0'
+        if torch.version.cuda is not None and TorchVersion(torch.version.cuda) >= required_cuda_version:
             cuda_compile_rule.append('  depfile = $out.d')
             cuda_compile_rule.append('  deps = gcc')
             # Note: non-system deps with nvcc are only supported