Home › Python

ImportError: DLL load failed while importing — how to fix

ImportError: DLL load failed while importing _multiarray_umath: The specified module could not be found.
ImportError: DLL load failed while importing QtCore: The specified procedure could not be found.
OSError: [WinError 126] The specified module could not be found. Error loading "...\torch\lib\c10.dll" or one of its dependencies.

Quick fix: Install the latest Microsoft Visual C++ Redistributable (x64), then reinstall the package in a clean virtual environment: python -m venv .venv, .venv\Scripts\activate, pip install --force-reinstall --no-cache-dir <package>. Don't mix conda and pip installs of the same package.

What it means

Python found the module (so it is not ModuleNotFoundError), but Windows refused to load the native .pyd file behind it. The message "The specified module could not be found" is misleading: the .pyd itself exists, but a DLL it depends on is missing or incompatible. "The specified procedure could not be found" means a DLL was found but it is the wrong version.

Typical victims: numpy, scipy, pandas, torch, PyQt5/PySide, opencv (cv2), psycopg2, lxml.

Cause 1: Visual C++ runtime missing

Most wheels are built with MSVC and need vcruntime140.dll, msvcp140.dll and friends. Fresh Windows installs and slim VMs often lack them.

Fix: install the latest "Microsoft Visual C++ Redistributable for Visual Studio 2015-2022" (x64) from Microsoft's site, or:

winget install Microsoft.VCRedist.2015+.x64

Restart the terminal and try again. For PyTorch this fixes most WinError 126 ... c10.dll errors.

Cause 2: broken or mixed installation

Upgrading one package with pip while another was installed by conda (or a different pip) leaves DLLs from different builds side by side.

Fix: reinstall cleanly, ideally in a fresh venv:

python -m venv .venv
.venv\Scripts\activate
pip install --upgrade pip
pip install --force-reinstall --no-cache-dir numpy

In conda, install with one tool only:

conda install numpy pandas      # not pip for the same packages
conda update --all

If you must use pip inside conda, install the native packages with conda first and only pure-Python ones with pip.

Cause 3: wrong architecture or Python version

A 32-bit Python cannot load 64-bit DLLs and vice versa, and wheels are built per Python minor version. Check:

python -c "import sys, platform; print(sys.version); print(platform.architecture())"
pip debug --verbose | findstr cp3

Fix: use 64-bit Python and a package version that ships wheels for your Python (very new Python releases often don't have them yet; use the previous minor version).

Cause 4: a conflicting DLL earlier on PATH

Another program (an old Qt, Anaconda, Intel MKL, a GPU toolkit) puts its own copy of a DLL on PATH, and Windows loads that one instead.

Fix: run from a clean shell with a minimal PATH, or find the conflict:

where Qt5Core.dll
where libiomp5md.dll

Remove the stray directory from PATH. Since Python 3.8, PATH is no longer searched for extension dependencies by default; packages call os.add_dll_directory() instead, so old advice to "add the folder to PATH" often doesn't help. In your own code:

import os
os.add_dll_directory(r"C:\path\to\dlls")
import mymodule

Cause 5: CPU or GPU runtime issues

To see exactly which dependency is missing, run the "Dependencies" tool (open-source successor of Dependency Walker) on the .pyd file named in the traceback.

Last updated 2026-11-08.