GPU observability setup

GPU Observability provides comprehensive hardware-level monitoring for GPU infrastructure used in AI workloads, essential for ensuring optimal performance and preventing hardware issues.

Install OpenLIT SDK

Install the OpenLIT SDK in your Python environment:

pip install openlit

Configure OTEL environment variables

Get your Grafana Cloud OTEL credentials

If you haven’t already obtained your OTEL credentials during the main AI Observability setup, follow these steps:

Sign in to Grafana Cloud and go to the Grafana Cloud Portal
Select your organization if you have access to multiple
Click your stack from the sidebar or main stack list
Under Manage your stack, click the Configure button in the OpenTelemetry section
Scroll down to the Password / API Token section and click Generate now (if you don’t have a token)
Enter a name for the token and click Create token
Click Close - you don’t need to copy the token manually
Scroll down and copy the OTEL_EXPORTER_OTLP_ENDPOINT and OTEL_EXPORTER_OTLP_HEADERS values from the Environment variables section

Set the environment variables

Set up the OpenTelemetry endpoints using the values you copied:

export OTEL_EXPORTER_OTLP_ENDPOINT="<YOUR_GRAFANA_OTEL_GATEWAY_URL>"
export OTEL_EXPORTER_OTLP_HEADERS="<YOUR_GRAFANA_OTEL_GATEWAY_AUTH>"

Replace the values with those you copied:

Replace <YOUR_GRAFANA_OTEL_GATEWAY_URL> with the OTEL_EXPORTER_OTLP_ENDPOINT value
Example: https://otlp-gateway-<ZONE>.grafana.net/otlp
Replace <YOUR_GRAFANA_OTEL_GATEWAY_AUTH> with the OTEL_EXPORTER_OTLP_HEADERS value
Example: Authorization=Basic%20<BASE64 ENCODED INSTANCE ID AND API TOKEN>

Initialize OpenLIT in your application

Add the following to your application code to enable GPU monitoring:

import openlit

openlit.init(collect_gpu_stats=True)

The OpenLIT SDK automatically uses the OTEL environment variables to send telemetry data to your Grafana Cloud instance. The collect_gpu_stats=True parameter enables GPU performance monitoring.

Alternative: GPU collector deployment

For containerized environments, you can use the dedicated GPU collector instead of the SDK approach:

Pull the GPU collector Docker image

docker pull ghcr.io/openlit/otel-gpu-collector:latest

Run the GPU collector container

docker run --gpus all \
    -e GPU_APPLICATION_NAME='my-ai-app' \
    -e GPU_ENVIRONMENT='production' \
    -e OTEL_EXPORTER_OTLP_ENDPOINT="<YOUR_GRAFANA_OTEL_GATEWAY_URL>" \
    -e OTEL_EXPORTER_OTLP_HEADERS="<YOUR_GRAFANA_OTEL_GATEWAY_AUTH>" \
    ghcr.io/openlit/otel-gpu-collector:latest

Use the same OTEL environment variable values you copied from Grafana Cloud Portal.

This approach is useful when you want to monitor GPU metrics system-wide rather than from within specific applications.

Visualize and analyze

With the GPU Observability data now being collected and sent to Grafana Cloud, the next step is to visualize and analyze this data to get insights into your GPU utilization, performance, and identify areas of improvement.

Navigate to the GPU Observability dashboard in your Grafana Cloud instance to start exploring. The dashboard provides: