Clears all KV / prefix cache tiers (GPU, CPU, and FS offload) on every vLLM pod in an llm-d deployment for a clean state — including both roles of a disaggregated prefill/decode setup. Use when the user wants to flush the cache, reset vLLM state, or start fresh before a test run — even if they don't say "cache" explicitly. Prefers the /reset_prefix_cache API; falls back to pod restart.
Configure an EPP EndpointPickerConfig for an llm-d deployment based on workload + SLA inputs, then optionally deploy it to the user's cluster. Activate when the user asks about configuring llm-d, setting up the inference scheduler, tuning EPP plugins, choosing what config to use for a model or workload, or deploying llm-d.
Configure and deploy an llm-d stack on an existing Kubernetes and OpenShift cluster using Well-Lit Path guides. Use this skill when users want to deploy, verify, or customize llm-d.
Compare the benchmark performance of two llm-d stack configurations end-to-end. For each configuration, deploys the stack, runs a benchmark, tears down, then generates a side-by-side results comparison. One configuration may already have been benchmarked previously — in that case, only the new configuration is deployed and benchmarked, and its results are compared against the existing ones. Use this skill whenever the user wants to A/B test llm-d configurations, compare llm-d config to baseline with no llm-d Inference Scheduler, compare models or serving strategies, evaluate the effect of a configuration change, run two benchmarks back-to-back for comparison, or compare a new run against a previous one — even if they say "which is faster", "test two setups", "compare these configs", "compare against my last run", or don't use "A/B" or "benchmark" terminology explicitly.
Benchmark an already-deployed llm-d stack using the llmdbenchmark CLI. Use this skill when the user wants to run a benchmark workload against a deployed llm-d stack — whether following a named guide (e.g. optimized-baseline, pd-disaggregation) or using a custom workload profile against any inference endpoint. Activate for requests to measure performance (TTFT, throughput, ITL), validate a deployment under load, or run inference-perf/guidellm/vllm-benchmark — even if they don't use the word "benchmark" explicitly.
Configure and deploy Workload Variant Autoscaler (WVA) for llm-d inference deployments. Guides users through namespace selection, WVA repo location, configuration (with presets or custom values), deployment via Makefile + kubectl apply, and verification. Produces a reusable deployment script.
Tears down, removes, cleans up, or undeploys a deployed llm-d stack on Kubernetes. Use this skill whenever the user wants to remove or destroy an llm-d deployment, clean up a namespace after inference workloads, uninstall helm releases for llm-d, or free up cluster resources used by llm-d — even if they don't say "teardown" explicitly. Also handles partial cleanup (e.g., remove only the model service, keep the gateway). Supports both Helm-based and Kustomize-based deployments.