NVIDIA Triton Inference Server: monitor server and per-model metrics via the built-in Prometheus metrics endpoint

Steps

1. Start `tritonserver` normally — Prometheus metrics are exposed by default at `http://localhost:8002/metrics` as plain text; nothing is pushed to a remote server, you must scrape/pull it
2. View or scrape metrics with `curl localhost:8002/metrics`, or point a Prometheus scrape config at that endpoint
3. Adjust exposure with CLI flags: `--metrics-port` to change the port, `--metrics-address` to bind a specific address, `--allow-metrics=false` to disable all metrics, `--allow-gpu-metrics=false`/`--allow-cpu-metrics=false` to disable just GPU/CPU metrics, and `--metrics-interval-ms` to change how often per-interval metrics are polled
4. Track core per-model counters such as `nv_inference_request_success`/`_failure`, `nv_inference_count`, `nv_inference_exec_count`, `nv_inference_pending_request_count`, and latency counters like `nv_inference_request_duration_us` and `nv_inference_queue_duration_us`
5. Track resource gauges such as `nv_gpu_utilization`, `nv_gpu_power_usage`, `nv_gpu_memory_used_bytes` (collected via DCGM), and `nv_cpu_utilization`/`nv_cpu_memory_used_bytes`
6. Optionally enable richer latency distributions with `--metrics-config histogram_latencies=true` or `--metrics-config summary_latencies=true` (both experimental) for percentile-level detail beyond the default counters