Troubleshooting¶
Common problems and fixes.
pixi: command not found¶
Restart your terminal after installing Pixi.
If it still fails, check that Pixi's binary directory is in your shell PATH.
Environment seems broken¶
Rebuild from the lock file:
rm -rf .pixi
pixi install
If you intentionally want to re-resolve dependencies:
rm -rf .pixi pixi.lock
pixi install
Be careful: deleting pixi.lock can produce different dependency versions.
CUDA is not available¶
Check CUDA from the GPU environment:
pixi run -e gpu python -c 'import torch; print(torch.cuda.is_available()); print(torch.version.cuda)'
If this prints False, check:
- you are using the GPU environment,
- your machine has an NVIDIA GPU,
- the NVIDIA driver is installed,
- the driver is compatible with the CUDA-enabled PyTorch package.
pixi run evaluate cannot find checkpoint¶
Run training first:
pixi run train
Then evaluate:
pixi run evaluate
Or pass a checkpoint manually:
pixi run python evaluate.py --config configs/default.yaml --ckpt path/to/checkpoint.ckpt
Aim UI opens but no runs appear¶
Make sure you trained first:
pixi run train
Then start Aim:
pixi run aim-ui
Open:
http://127.0.0.1:43800
Look for the ml-template experiment.
Local files are not committed¶
This is expected for generated outputs under:
local/
The local/ directory is ignored by git to keep experiment outputs, checkpoints, figures, and temporary files out of version control.
Tests fail after modifying the template¶
Run:
pixi run pytest
Then check whether the failure is from:
- renamed modules,
- changed config keys,
- missing imports,
- removed example files,
- or changed expected output paths.
Update the smoke tests after major structural changes.