iket-profiling-skill

An Agent Skill for profiling inside CuTe DSL kernels with IKET
(In-Kernel Event Tracing) and the run-iket profiler.

This is not new work. The skill is a condensed summary of NVIDIA's official CUTLASS guide,
IKET Profiling,
written for coding agents. The guide is the source of truth; where they differ, follow the guide.
IKET is experimental, and its API and output may change.

Table of Contents

• Install
• Usage
• Maintainers
• Contributing
• License

Install

The skill is the skills/iket-profiling/ directory. Install it with
the skills CLI:

npx skills add humanfia/iket-profiling-skill

or copy the directory into your agent's skills directory, for example for Claude Code:

git clone https://github.com/humanfia/iket-profiling-skill.git
cp -r iket-profiling-skill/skills/iket-profiling ~/.claude/skills/

The directory name must stay iket-profiling, the skill's name.

Usage

The agent loads the skill when you develop or tune a CuTe DSL (cutlass.cute) kernel and ask
where its time goes inside the kernel: phase timelines, TMA/MMA activity, or pipeline waits. It
covers the cutlass.cute.experimental.iket API, the run-iket workflow, instrumentation rules,
the Perfetto and JSON output, limitations, and troubleshooting. You need an nvidia-cutlass-dsl
installation that includes run-iket and an SM90 or newer GPU.

Maintainers

@DongyunZou

Contributing

Issues and pull requests are welcome; see the
contributing guide. Every
statement in the skill must be backed by the official guide. Check changes with
skills-ref:

uvx --from "git+https://github.com/agentskills/agentskills#subdirectory=skills-ref" \
  skills-ref validate skills/iket-profiling

License

Apache-2.0 © Humanfia, for the text of this repository. The skill summarises and
adapts NVIDIA's CUTLASS documentation, which is © NVIDIA Corporation under the BSD-3-Clause
license; see NOTICE. IKET, CuTe DSL, and CUTLASS are NVIDIA projects.