Bug: Detectron2 fails to build with system CUDA 13 + PyTorch 2.9.1

Instructions To Reproduce:

1. Full runnable code or full changes you made:

# No source modifications.
# Fresh clone of detectron2.

git rev-parse HEAD; git diff

<no changes>

2. What exact command you run:

uv venv --python=3.12 .venv
source .venv/bin/activate

uv pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128

uv pip install -e detectron2 --no-build-isolation

3. Full logs or other relevant observations:

/usr/local/cuda/include/crt/host_config.h:143:2: error:
#error -- unsupported GNU version! gcc versions later than 12 are not supported!
nvcc fatal   : Host compiler targets unsupported GNU version

W1205 torch/utils/cpp_extension.py: MIG mismatch:
"The detected CUDA version (12.2) has a minor version mismatch with the version that was used to compile PyTorch (12.8)."

error: command '/usr/local/cuda/bin/nvcc' failed with exit code 1

Full trace excerpt:

In file included from /usr/local/cuda/include/cuda_runtime.h:82:
/usr/local/cuda/include/crt/host_config.h:143:2:
#error -- unsupported GNU version! gcc versions later than 12 are not supported!
...
The detected CUDA version (12.2) has a minor version mismatch with the version that was used to compile PyTorch (12.8)
...
distutils.errors.CompileError: command '/usr/local/cuda/bin/nvcc' failed with exit code 1

Additional observation:

My system CUDA toolkit is version 13.x, but PyTorch wheels only ship with CUDA 12.8, which creates a mismatch. Detectron2 attempts to compile with system nvcc (CUDA 13) which rejects gcc 13.

4. Steps simplified to minimal reproducible example:

No custom datasets or external resources required. Just:

git clone https://github.com/facebookresearch/detectron2
cd detectron2
uv venv --python=3.12 .venv
source .venv/bin/activate
uv pip install torch==2.9.1+cu128 torchvision==0.24.1+cu128 --index-url https://download.pytorch.org/whl/cu128
uv pip install -e . --no-build-isolation

This fails consistently on any clean environment with CUDA 13 installed system-wide.

Expected behavior:

Detectron2 should compile its CUDA/C++ extensions successfully, or at least detect and warn when the system CUDA toolkit version (13) is not supported.

Given that PyTorch 2.9.1 supports CUDA 12.8, I expect Detectron2 to either:

• build correctly with CUDA 12.8 (without using system CUDA 13), or
• provide instructions for using PyTorch’s CUDA toolkit instead of system nvcc.

Environment:

(When running the recommended script, replace <output> below with your actual output:)

wget -nc -q https://github.com/facebookresearch/detectron2/raw/main/detectron2/utils/collect_env.py && python collect_env.py

My known relevant environment details:

• GPU: NVIDIA RTX 4090
• System CUDA: 13.x (/usr/local/cuda)
• PyTorch: 2.9.1+cu128
• TorchVision: 0.24.1+cu128
• Python: 3.12.11
• Compiler: gcc 13.x
• OS: Arch Linux
• Install tool: uv

The script output will confirm this mismatch:

• PyTorch CUDA = 12.8
• System CUDA = 13.x
• gcc > 12 (nvcc from CUDA 13 rejects this)

Additional Notes

I cannot downgrade my system CUDA installation (CUDA 13 is required for other software), and this appears to make Detectron2 impossible to build, because:

• Detectron2 does not yet support CUDA 13
• CUDA 13’s nvcc rejects GCC 13
• PyTorch does not ship CUDA 13 wheels
• Building against system CUDA (13) while PyTorch is built for CUDA 12.8 causes ABI mismatch and build failure