Project overview
Use the playbook directly or install its Claude skill to investigate slow training, low utilization, data-loading stalls, out-of-memory failures, weak multi-GPU scaling, and LLM serving latency or throughput problems. Its workflow starts with measurement, classifies the limiting resource, applies the smallest relevant change, and verifies the result. Separate guides cover training, inference, distributed systems, kernels, profilers, and recurring performance anti-patterns in code review.
Repository facts
- Primary language
- Shell
- License
- MIT
- Repository updated
- Apr 11, 2026
- Default branch
- main
Resource types
General skill
Use cases
Testing and debugging
Platforms
Claude Code, Codex, and more
Capabilities
Verification and evals