Deployment documentation

For GKE, EKS, AKS, and self-managed K8s.

helm repo add pebble https://charts.gopebble.com
helm install pebble-sonar pebble/sonar \
  --set gpuVendor=nvidia \
  --set inferenceEngine=vllm

1. Add the Pebble Helm repository and update.
2. Install the Sonar chart, pointing at your inference engine (vLLM, SGLang, or TensorRT-LLM).
3. Pebble auto-discovers GPU nodes via the NVIDIA / ROCm device plugins.
4. Verify with kubectl get pods -n pebble-system; metrics show in Grafana within ~60 seconds.

For on-prem HPC clusters and bare-metal AI factories.

ansible-galaxy collection install pebble.flex
ansible-playbook -i hosts.ini \
  pebble.flex.deploy.yml

1. Install the Ansible collection from Galaxy.
2. Provide your Slurm controller and compute node inventory.
3. The playbook deploys Pebble agents and the SLA-cliff profiler on every GPU node.
4. First curtailment signal arrives within minutes of activation.

For local benchmarks and single-node validation.

docker run --gpus all \
  -e PEBBLE_MODE=sonar \
  -v /var/run/dcgm:/dcgm \
  gopebble/agent:latest

1. Pull the Pebble agent image.
2. Mount your DCGM (NVIDIA) or rocm-smi (AMD) socket so the agent can read GPU telemetry.
3. Run for ~10 minutes against a representative workload to see your efficiency-knee profile.