Enforce a finite state machine loop for NVIDIA GPU kernel tuning.
npx skills add https://github.com/codes1gn/CroqTuner --skill fsm-engine
Please help me install this Agent Skill. Skill: fsm-engine Source: https://github.com/codes1gn/CroqTuner/tree/main/.claude/skills/fsm-engine Command: npx skills add https://github.com/codes1gn/CroqTuner --skill fsm-engine
This Skill solves the challenge of running long, error-prone GPU kernel optimization experiments by enforcing a finite state machine workflow with persistent state, validation gates, and autonomous resume support.
Use the fsm-engine skill to start a CroqTuner GPU kernel optimization run from the best existing kernel for the specified shape and data type.
High-intent search queries and answers about installing and using this skill.
A finite state machine workflow automates GPU kernel optimization by enforcing structured states like profiling and measurement, using persistent state files and checkpoints to preserve experiment progress reliably across multiple sessions.
Managing sparse GEMM tuning requires validation scripts, idea deduplication, and checkpointing. These guardrails prevent skipped steps and repeated experiments by strictly enforcing a structured finite state machine optimization loop.
Yes, you can resume NVIDIA GPU kernel profiling workflows after interruptions using autonomous resume support. The finite state machine relies on persistent state files and compaction summaries to recover and continue tuning iterations reliably.
Yes, you need jq and git installed. These dependencies are required to execute the validation scripts, manage state files, and track persistent experiment data during your automated CUDA kernel tuning workflows.
A finite state machine controls iterative code generation by cycling through defined states: profiling, ideation, implementation, measurement, decision, and storage. This ensures benchmark-driven improvements are validated before advancing to the next optimization phase.
fla-orgProfile and benchmark NVIDIA GPU kernels in the FLA repository.
flashinfer-aiExtract GPU kernel schemas from SGLang models into deduplicated Definition JSON files.
m0atAnalyze CUDA kernel performance bottlenecks with Nsight Systems and Nsight Compute.
goupup-aiAutomate iterative parameter tuning for electronic design automation and computer architecture workflows.
