# Python inference

The two Python runtimes are isolated because they target different hardware and dependencies.

| Directory | Backend | Model files |
| --- | --- | --- |
| `ax650/` | PyAXEngine on AX650 / NPU3 | Packaged `*.axmodel` files |
| `pytorch/` | PyTorch on CPU, CUDA, or MPS | Original upstream FP checkpoints |

See the repository root `README.md` for installation and run commands.