Observability for NVIDIA NIM for LLMs

NIM provides Prometheus metrics indicating request statistics. You can use these metrics to create dashboards with Grafana. By default, these metrics are available at http://localhost:8000/v1/metrics.

The following table describes the available metrics.

Category	Metric	Metric Name	Description	Granularity	Frequency
KV Cache	GPU Cache Usage	gpu_cache_usage_perc	GPU KV-cache usage. 1 means 100 percent usage	Per model	Per iteration
Count	Running Count	num_requests_running	Number of requests currently running on GPU	Per model	Per iteration
	Waiting Count	num_requests_waiting	Number of requests waiting to be processed	Per model	Per iteration
	Max Request Count	num_request_max	Max number of requests that can be run concurrently by the model	Per model	Per iteration
	Total Prompt Token Count	prompt_tokens_total	Number of prefill tokens processed	Per model	Per iteration
	Total Generation Token Count	generation_tokens_total	Number of generation tokens processed	Per model	Per iteration
Latency	Time to First Token	time_to_first_token_seconds	Histogram of time to first token in seconds	Per model	Per request
	Time per Output Token	time_per_output_token_seconds	Histogram of time per output token in seconds	Per model	Per request
	End to End	e2e_request_latency_seconds	Histogram of end to end request latency in seconds	Per model	Per request
Count	Prompt Token Count	request_prompt_tokens	Histogram of number of prefill tokens processed	Per model	Per request
	Generation Token Count	request_generation_tokens	Histogram of number of generation tokens processed	Per model	Per request
	Finished Request Count	request_finish_total	Number of finished requests, with label indicating finish reason	Per model	Per request
	Success Request Count	request_success_total	Number of successful requests, requests with finish reason “stop” or “length” are counted	Per model	Per request
	Failure Request Count	request_failure_total	Number of failed requests, requests with other finish reason are counted	Per model	Per request

Prometheus

To install Prometheus for scraping metrics from NIM, download the latest Prometheus version appropriate for your system.

wget https://github.com/prometheus/prometheus/releases/download/v2.52.0/prometheus-2.52.0.linux-amd64.tar.gz
tar -xvzf prometheus-2.52.0.linux-amd64.tar.gz
cd prometheus-2.52.0.linux-amd64/

Edit the Prometheus configuration file to scrape from the NIM endpoint. Make sure the targets field point to localhost:8000

vi prometheus.yml

# A scrape configuration containing exactly one endpoint to scrape:
# Here it's Prometheus itself.
scrape_configs:
  # The job name is added as a label `job=<job_name>` to any timeseries scraped from this config.
  - job_name: "prometheus"

    # In Prometheus 2.53, metrics_path defaults to '/metrics'
    # Previous versions use '/v1/metrics'.
    # scheme defaults to 'http'.

    static_configs:
      - targets: ["localhost:8000"]

Next, run the Prometheus server: ./prometheus --config.file=./prometheus.yml

Use a browser to verify that the NIM target was detected by the Prometheus server: http://localhost:9090/targets?search=. You can also click on the NIM target URL link to explore generated metrics.

Grafana

You can use Grafana for visualizing NIM metrics. Install the latest Grafana version for your system.

wget https://dl.grafana.com/oss/release/grafana-11.0.0.linux-amd64.tar.gz
tar -zxvf grafana-11.0.0.linux-amd64.tar.gz

Run the Grafana server:

cd grafana-v11.0.0/
./bin/grafana-server

To access the Grafana dashboard, point your browser to http://localhost:3000. Log in using the defaults:

username: admin 
password: admin

First, configure the data source in Grafana. Click Data Source, select Prometheus, and set the URL to localhost:9090. After saving the configuration, you should see a success message. You can now create a dashboard using NIM metrics or import this example dashboard.