Miniconda+Python3.8环境下Tensorflow缺失msvcp140_1.dll问题及解决方法

What I Ran Into #

I set up a TensorFlow environment using Miniconda with this command:

conda create -n tf tensorflow

But when I tried running my code in VS Code using the tf environment's interpreter, I kept getting an error saying msvcp140_1.dll couldn't be found. TensorFlow's error message told me to install the Microsoft C++ Redistributable for Visual Studio 2015, 2017 and 2019, so I reinstalled the x64 version three times. I even confirmed the DLL existed in multiple locations:

• C:\Users\myusername\miniconda3\envs\tf\msvcp140_1.dll
• C:\Users\myusername\miniconda3\envs\tf\Library\bin\msvcp140_1.dll
• C:\Windows\System32\msvcp140_1.dll
• C:\Users\myusername\miniconda3\Library\bin\msvcp140_1.dll
• C:\Users\myusername\miniconda3\msvcp140_1.dll

Most online fixes just said to install that redistributable package, but that didn't work for me at all. I spent hours digging through forums until I found a solution from Eryk Sun's comments that finally fixed the problem.

The Fix That Worked #

I modified TensorFlow's self_check.py file to add winmode=0 to the ctypes.WinDLL call. Here's how:

1. Navigate to the self_check.py file in your Conda environment's TensorFlow folder. For me, that was:

   C:\Users\myusername\miniconda3\envs\tf\Lib\site-packages\tensorflow\python\platform\self_check.py
   

2. Find the line where ctypes.WinDLL is being used to load a DLL related to the redistributable check.
3. Update the line to include winmode=0:
   Before:

   ctypes.WinDLL(dll_name)
   

   After:

   ctypes.WinDLL(dll_name, winmode=0)
   

4. Save the file, then re-run your TensorFlow code. The error should be gone.

Why This Works (Thanks to Eryk Sun's Technical Explanation) #

Starting with Python 3.8, the DLL search behavior changed. The old way where ctypes.WinDLL(dll_name) used the system PATH to find DLLs is no longer the default. Now Python only looks in:

• The application directory (like your Conda env's sys.prefix folder)
• The System32 directory
• Paths you've explicitly added using os.add_dll_directory()

TensorFlow's code was still relying on the old PATH-based search. Adding winmode=0 forces the WinDLL call to use the legacy search behavior, which picks up the DLL even if it's in a path that isn't in the process's official DLL search list.

I can't say if this is the perfect long-term solution, but it worked immediately for me when nothing else did.

内容的提问来源于stack exchange，提问作者Erik