Device Management Overview

EagleEye does not wrap accelerators in a device abstraction that executes models. Instead it keeps three separate objects, all created once in MainBackend.__init__ and injected into operations that ask for them by constructor parameter name:

Object	File	Role
DeviceRegistry	src/utils/device_registry.py	Immutable inventory of inference devices discovered at startup
ModelLibrary	src/utils/model_library.py	Managed model metadata and artifacts under files/models/
Mx3RuntimeCoordinator	src/utils/mx3_runtime.py	Owner of shared per-device MemryX MX3 runtimes and stream bindings

An operation that runs inference typically takes a model_id and a device_id parameter, resolves the artifact with model_library.resolve_artifact(model_id, device_id), and loads it itself with the appropriate framework (Ultralytics/PyTorch, ONNX Runtime, TensorRT, or the MX3 coordinator).

Device ID formats

Device	Canonical ID	Example
CPU	cpu	cpu
NVIDIA CUDA GPU	cuda:<index>	cuda:0, cuda:1
MemryX MX3	mx3:<index>	mx3:0, mx3:1

IDs are exact; aliases are not accepted. CUDA indices follow torch.cuda enumeration order. MX3 indices come from the /dev/memxN node number.

Lifecycle

1. DeviceRegistry.discover(logger=...) runs once in MainBackend.__init__. It always adds cpu, adds one entry per CUDA device if torch imports and reports CUDA available, and adds one entry per /dev/memx[0-9]* node on POSIX systems.
2. The registry, model library, and MX3 coordinator are passed to EagleEyeInterface (registry and library only) and to generate_all_pipelines(...).
3. Pipeline construction injects device_registry, model_library, and mx3_coordinator into any operation whose __init__ declares those parameter names.
4. Operations validate their configured device_id through DeviceRegistry.get(device_id), which raises DeviceNotFoundError for unknown IDs.

There is no runtime registration or removal API: the inventory is fixed for the life of the process, and adding hardware requires a backend restart.

When to add a new device

Supporting new accelerator hardware means extending DeviceRegistry.discover with a new canonical ID prefix and teaching ModelLibrary.resolve_artifact which artifact slot that prefix can use. See New Device.