atrex-kernel-agent

End-to-end GPU kernel implementation and optimization router. Use this skill to turn PyTorch logic into a high-performance kernel or to systematically optimize an existing kernel. It initializes the local knowledge base, identifies the current phase, and routes work to baseline implementation, bottleneck analysis, and profile-driven optimization sub-skills.

Learn the target framework from gpu-wiki and implement a baseline GPU kernel. Use this skill to understand compute semantics, determine the target platform and framework, search reference implementations, and produce a correct V0 baseline with performance records for later profile-driven optimization.

Helper skill for GPU kernel bottleneck analysis. It provides Roofline analysis, same-size bandwidth baseline measurement, TFLOPS and bandwidth utilization calculation, and profiling evidence extraction. This skill is not a standalone stage; its capabilities are reused by Step 0 and by the first Stage 2 iteration.

Final output contract for GPU kernel optimization tasks. Use this skill when a verified implementation must be packaged as generated_kernel.py for hidden evaluator import, especially when a stop hook blocks because the final candidate file is missing or non-compliant.

Partial-restart workflow for GPU kernel optimization when no new optimization direction is available. The main agent masks half of the optimization memory, then launches a fresh subagent that reads README.md and unmasked memory to extract experience, takes the latest kernel.py as the starting point, and continues optimization through the gpu-kernel-profile-optimizer workflow.

Profile-driven GPU kernel iterative optimization skill. Use this skill to run a closed loop in a temporary git workspace: profile evidence extraction, evidence-driven search and planning, single-category optimization, validation, memory update, and git commit.