# 🔨 Build Notes

- Enable exactly one GPU back-end that matches your device (`CUBLAS`, `METAL`, `OPENCL`, …).  
- For very large models, more VRAM helps, but OptiML’s hybrid placement reduces the requirement.  
- Use quantization to lower memory and often improve speed on PC-class hardware.  
- Ensure release builds (`-DCMAKE_BUILD_TYPE=Release`) for best performance.