vLLM-HUST runtime metrics collection and reporting adapter

[kms12425-ctrl]

Background

Real-time execution state is required to drive memory-aware scheduling and model residency decisions.

Scope

• Export runtime metrics from vLLM-HUST instances.
• Report both periodic heartbeat metrics and event-driven memory events.
• Surface model-weight residency and KV cache behavior needed for multi-model control.

Deliverables

• Metrics collection component.
• Reporting protocol and metrics dictionary.
• Sample dashboards or query examples.

Acceptance Criteria

• Collect: free_vram_bytes, reserved_vram_bytes, model_weight_bytes, model_residency_state, kv_cache_bytes, kv_cache_utilization, prefix_cache_hit_rate, queue_depth, inflight_requests, expected_ttft_ms, expected_tpot_ms, model_load_state.
• Report memory allocation/release/reuse events for model weights and KV cache.
• Support both heartbeat and event-driven reporting paths.

Dependencies

• Epic: Epic: SLO-aware LLM inference service system on SAGE #1506

Notes

Connects execution-plane memory state to control-plane scheduling.