Environment
Validated Jittor setup
| Component | Version |
|---|---|
| Python | 3.10 |
| Jittor | 1.3.8.5 |
| NumPy | 1.26.4 |
| CUDA toolchain | Jittor CUDA 11.2 + cuDNN 8 |
| C++ compiler | g++-10 |
Install the Python dependencies and expose the repository package:
pip install -r requirements.txt
export cc_path=/usr/bin/g++-10
export PYTHONPATH="$PWD:$PWD/python"
Jittor compiles operators on first use, so the first model construction can take several minutes.
Compatibility notes
NumPy
Keep NumPy at 1.26.4. Jittor 1.3.8.5 combined with NumPy 2.x can produce incorrect values for operations consuming NumPy-backed arrays without raising an exception. Verify the installation with:
python - <<'PY'
import jittor as jt
x = jt.float32([1, 2, 3])
assert (x + x).numpy().tolist() == [2.0, 4.0, 6.0]
print('Jittor array check passed')
PY
Compiler
Jittor's CUDA 11.2 frontend is incompatible with newer system compiler headers
on some Linux distributions. The validated toolchain uses g++-10 selected by
the lowercase cc_path environment variable.
Reference environment
Regenerating PyTorch goldens or using the reference DOTA merge metric requires the original Point2RBox-v3 environment: PyTorch 2.2, torchvision 0.17, mmengine 0.10.7, mmcv 2.2.0, mmdet 3.3.0 and the reference mmrotate package. Regular Jittor training and inference do not require this second environment.