feat(backends): add NanoDeploy backend with dlslime-ctrl discovery

[JimyMa]

Summary

• Integrate NanoDeploy's single-process OpenAI server (nanodeploy serve) as a first-class DLRouter backend (--backend nanodeploy).
• Add BackendType.NANODEPLOY and the nanoctrl service-discovery mode that polls a dlslime-ctrl entity registry for nanodeploy nodes and reconciles their HTTP endpoints into the NodeManager (served-model-name, model-path, and basename aliases).
• Auto-discovery activates in hybrid serving when --ctrl_address is set; manual POST /nodes/add still works otherwise.
• Docs: README updated with a supported-backend row, a dedicated NanoDeploy + dlslime-ctrl quick start, and a request example.

Running NanoDeploy with DLRouter

1. Start the dlslime-ctrl control plane (only needed for auto-discovery)

dlslime-ctrl server --redis-url redis://127.0.0.1:6379

2. Start the NanoDeploy OpenAI server

# inside the nanodeploy conda env
nanodeploy serve /path/to/Qwen3-0.6B \
  --host 0.0.0.0 --port 8100 \
  --served-model-name Qwen3-0.6B \
  --ctrl_address 127.0.0.1:4479

Notes:

• The positional argument is the model path (you can also use --model /path/to/...).
• --served-model-name is the public model id; if omitted it defaults to the basename of the model path.
• --ctrl_address enables self-registration + heartbeat to dlslime-ctrl. Omit it to run as a standalone HTTP server.
• All other Config fields (--ray_address, --tp, etc.) share the same names/semantics as engine_server.py. --host/--port bind the uvicorn HTTP API.

3. Call NanoDeploy directly (bypass DLRouter, verify the server itself)

curl http://localhost:8100/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-0.6B","messages":[{"role":"user","content":"Hello"}]}'

Other endpoints:

curl http://localhost:8100/health        # health check
curl http://localhost:8100/v1/models     # served-name / path / basename are all aliases

# /v1/completions (text completion)
curl http://localhost:8100/v1/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-0.6B","prompt":"Once upon a time","max_tokens":64}'

# streaming
curl -N http://localhost:8100/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-0.6B","messages":[{"role":"user","content":"Hello"}],"stream":true}'

4. Call through DLRouter (end-to-end)

# DLRouter auto-discovers NanoDeploy nodes from dlslime-ctrl
python -m dlrouter \
  --backend nanodeploy \
  --serving_strategy hybrid \
  --ctrl_address 127.0.0.1:4479

# Request hits port 8000 (DLRouter), which forwards to 8100 (NanoDeploy)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"Qwen3-0.6B","messages":[{"role":"user","content":"Hello"}]}'

Without dlslime-ctrl, drop --ctrl_address on the DLRouter side and register the node manually:

curl -X POST http://localhost:8000/nodes/add \
  -H "Content-Type: application/json" \
  -d '{"url":"http://127.0.0.1:8100"}'

Test plan

• pytest tests/core/test_nanoctrl_discovery.py tests/backends/test_backend_contracts.py
• End-to-end: nanodeploy serve + dlslime-ctrl + python -m dlrouter --backend nanodeploy --serving_strategy hybrid --ctrl_address 127.0.0.1:4479, then a /v1/chat/completions curl (verified working manually).