Troubleshooting Common Issues
GPU Not Detected
If you see CPU-only mode:
- Verify PyTorch with Metal support: python -c “import torch; print(torch.backends.mps.is_available())”
- Ensure you are on Apple Silicon
- Check macOS version is 13+
Out of Memory
- Reduce context length in settings
- Use smaller models (7B instead of 70B)
- Clear model cache: rm -rf ~/.cache/
Slow Inference
- Check GPU utilization: Activity Monitor > GPU
- Use quantised models (Q4_K_M)
- Reduce batch size
Connection Issues
If nodes cannot communicate:
- Check Thunderbolt bridging: System Preferences > Network
- Verify firewall rules
- Test with ping: ping 172.16.0.x