iket-profiling-skill

This is NOT new work. This repo is just a condensed summary of NVIDIA's official CUTLASS documentation on IKET profiling for CuTe DSL kernels, repackaged as an agent skill. All technical content comes from the official docs — treat them as the source of truth:

• Official guide: https://docs.nvidia.com/cutlass/latest/media/docs/pythonDSL/cute_dsl_general/iket_profiling.html
• Official example: examples/python/CuTeDSL/dsl_tutorials/fp16_gemm_4_iket.py in the CUTLASS repo

What it is

A single SKILL.md in the open Agent Skills format (YAML frontmatter + markdown). It activates when an agent is developing a CuTe DSL (cutlass.cute) kernel and tuning its performance, and covers: the IKET API (mark / range_push·range_pop / range_start·range_end / sentinel_token), the run-iket profiler workflow, instrumentation rules, common patterns (async issue vs. wait timing, cross-iteration ranges), output formats (Perfetto / JSON), limitations, and troubleshooting — all per the official guide.

Install

Claude Code (personal, all projects):

git clone git@github.com:humanfia/iket-profiling-skill.git ~/.claude/skills/iket-profiling

or into a project: .claude/skills/iket-profiling/.

Codex CLI (versions supporting the Agent Skills standard):

git clone git@github.com:humanfia/iket-profiling-skill.git ~/.codex/skills/iket-profiling

If your agent doesn't support skills, just point it at SKILL.md (e.g. reference it from AGENTS.md).

Attribution

IKET, CuTe DSL, and CUTLASS are NVIDIA projects. This summary was last checked against the docs in August 2026; IKET is experimental and its API/output may change — always defer to the official documentation.