AI-Infra-Auto-Driven-SKILLS

Why we included this project

For AI infrastructure engineers, the value here is that an agent can actually do the work instead of producing generic advice. The repo packages operational knowledge as playbooks: fair benchmarks for SGLang, vLLM, TensorRT-LLM, and TokenSpeed, Day-0 support plans for new SGLang model architectures, and capacity estimates read from startup logs. Profiler triage keeps prefill and decode evidence separate, and the skills cover the kind of detail that usually lives in a senior engineer's head, like inspecting traces at layer and kernel level or estimating operator FLOPs and MFU. SGLang patch review is grounded in how maintainers actually discuss changes, and there's production incident triage from replays plus a diff-backed history of model-family optimization PRs kept next to the code that changed. Teams running LLM serving stacks can hand these playbooks to an agent and get auditable, reproducible work instead of confident-sounding guesses.