Auto-Discovery¶
Under Development
Peer-to-peer auto-discovery is currently under development. While local model discovery is functional, automatic discovery of other GO instances on the local network will be available in a future update.
In a distributed environment, manually configuring every single model or service would be tedious. Backend.AI GO utilizes an Auto-Discovery mechanism to simplify resource management.
Service Discovery within the Mesh¶
Once you establish a connection to a node (be it a Cluster or another GO instance), you don't need to manually register every model available on that node.
- Handshake: When a connection is added, Backend.AI GO performs an initial handshake.
- Catalog Exchange: The remote node sends a list of its currently active "backends" (loaded models, services).
- Dynamic Updates:
- If a new model is loaded on the remote cluster, your local client is notified (or discovers it on the next refresh).
- These remote models appear in your chat interface's model selector, grouped by the provider/node name.
Local Model Discovery¶
On your local machine, "Auto-Discovery" works by monitoring the file system and active processes:
- Folder Monitoring: When you download a model to the
models/directory, Backend.AI GO automatically detects the file type (GGUF, safetensors) and determines the appropriate runner (llama.cpp, vLLM, etc.). - Process Binding: When a model loads, it registers itself with the local Continuum Router. The Router then announces this capability to the frontend UI.
Troubleshooting Discovery¶
If remote models are not appearing:
- Check Network: Ensure the connection status in Settings > Mesh is green.
- Permissions: On a Cluster, ensure your Access Key has permission to view the specific Session or vFolder where the model is hosted.
- Refresh: Sometimes a manual refresh (or restarting the client) is needed to trigger a full catalog sync.