feat: backport monitoring metrics and fix stream label / serve_count bugs

[m199369309]

Summary

• Monitoring metrics backport (2.7.0 → 2.8.0): Restores 5 customized metric changes that were overwritten during the 2.8.0 upgrade:

  • model_last_load_duration_seconds Gauge with load timing in worker
  • _WORKER_ONLY_METRICS set + Supervisor-side deregister to avoid empty HELP/TYPE headers
  • image/audio/video engine/format label differentiation (empty labels instead of "unknown")
  • grafana_alert_datasource in UI config API + grafanaUtils.js URL params
  • @request_limit decorator on image_to_image, inpainting, ocr methods
  • Exception protection for record_metrics

• Stream label fix: request_limit decorator now recognizes xoscar IteratorWrapper as a stream type, fixing stream="true" label never being recorded for LLM requests

• serve_count double-decrement fix: Added stream type detection guard (isasyncgen/isgenerator/IteratorWrapper) to 3 stream_results() finally blocks, ensuring decrease_serve_count() is only called when the decorator has deferred the decrement to the caller. This prevents model_serve_count from going negative in two scenarios:

  • model.chat() fails before iterator is assigned (iterator is None)
  • model.chat() returns a non-stream value on stream=True path (iterator is not a generator/IteratorWrapper)

Test plan

• Deploy and verify model_last_load_duration_seconds is reported after model load
• Verify Supervisor /metrics endpoint no longer shows worker-only metric HELP/TYPE headers
• Send stream + non-stream LLM requests, confirm model_request_total{stream="true"} and stream="false" are both recorded correctly
• Verify model_serve_count stays ≥ 0 after failed stream requests
• Verify grafana_alert_datasource appears in /v1/cluster/ui_config response
• Verify image/audio model metrics show empty engine/format labels instead of "unknown"

[gemini-code-assist]

Code Review

This pull request enhances metrics management by introducing model load duration tracking and separating worker-only metrics. It also improves API robustness by adding request limits to image models and refining serve count handling in streaming endpoints. Additionally, the UI and Grafana utilities were updated to support alert datasources and time ranges. Review feedback identifies opportunities to further stabilize serve count decrements by aligning API stream detection with decorator logic and suggests using relative imports for consistency.

[qinxuye]

LGTM