What problem does it solve? GPU inference services on Kubernetes often scale on the wrong signals (CPU), thrash expensive GPU nodes, or leave pods Pending because the pod-scaling and node-scaling loops are misconfigured. This Skill designs both loops together so capacity follows demand without paying for idle accelerators. ## Core Features & Use Cases - Dual-trigger ScaledObjects: Authors KEDA ScaledObjects combining a demand metric (RPS, queue depth) with an experience or GPU metric (p95 latency, TTFT, DCGM_FI_DEV_GPU_UTIL), including scale-to-zero and asymmetric scale-up/scale-down behavior. - GPU node-pool policy: Designs tainted dedicated GPU pools with aggregate GPU-count limits, WhenEmpty consolidation warm pools, startup taints, and cold-start engineering via image pre-caching. - YAML review and validation: Reviews existing ScaledObject/HPA configurations for GPU services and validates the full chain from metric trigger to node provisioning. - Use Case: Your LLM inference pods scale up but new replicas sit Pending for minutes while GPU nodes provision. Use this Skill to tune stabilization windows, configure a warm pool with consolidateAfter, and add a queue-depth trigger so scale-up completes before users notice. ## Quick Start Use the gpu-autoscaling-engineer skill to author a KEDA ScaledObject with scale-to-zero and dual Prometheus triggers for my llm-inference deployment.
