Cuda version mismatch

[PremHCLTech]

❓ Questions and Help

General questions about detectron2.

The detected CUDA version (12.1) mismatches the version that was used to compile
PyTorch (11.7). Please make sure to use the same CUDA versions. while runnning

python3 setup.py build develop

NOTE:

If you met any unexpected issue when using detectron2 and wish to know why,
please use the "Unexpected Problems / Bugs" issue template.

[tharun571]

The CUDA version mismatch error happens because your system has CUDA 12.1 installed, but PyTorch was compiled against CUDA 11.7.

The issue:

The detected CUDA version (12.1) mismatches the version that was used to compile PyTorch (11.7)

This breaks during python3 setup.py build develop when detectron2 tries to compile CUDA extensions.

Solution approaches:

1. Install PyTorch for CUDA 12.1 (recommended):

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Then rebuild detectron2:

python3 setup.py build develop

2. Or downgrade CUDA to 11.7 (if you need CUDA 11.7 specifically):

• Uninstall CUDA 12.1
• Install CUDA 11.7
• Reinstall PyTorch for CUDA 11.7

3. Verify compatibility before installing:
   I built env-doctor to catch these mismatches early:

pip install env-doctor
env-doctor check

It will tell you:

• Your GPU driver's maximum CUDA support
• Which PyTorch wheel matches your system
• The exact pip install command to use

Then you can install the right PyTorch version before building detectron2, avoiding this error entirely.

Full disclosure: I'm the author. I built this specifically for cases like yours where framework extensions fail to compile due to CUDA version mismatches.