[Agents] Add Triton kernel-writing skill

[WoosukKwon]

Summary

• Add a concise, repository-local Triton kernel-writing skill.
• Cover generated-code inspection, tuning and specialization choices, launch-grid limits, 64-bit offsets, read-after-write synchronization, and decode-versus-prefill tiling.
• Require representative shape sweeps and numerically difficult correctness inputs.

Duplicate-work check

Open PR searches for Triton kernel skill, agent skill Triton, kernel-writing, triton-kernel-writing, and Triton Kernel Writing found no PR addressing this change. There is no linked issue.

This is materially different from closed PR #55009. That PR imported and adapted a much larger TensorRT-LLM skill under kernel-triton-writing, including provenance and reference documents. This PR instead adds a clean, focused, vLLM-specific skill under triton-kernel-writing with only the guidance requested here and no imported content.

Testing

• .venv/bin/python /home/woosuk/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/triton-kernel-writing — passed.
• .venv/bin/pre-commit run --files .agents/skills/triton-kernel-writing/SKILL.md .agents/skills/triton-kernel-writing/agents/openai.yaml — all applicable hooks passed.

Model evaluation

Not applicable. This adds agent guidance only and does not change model output, accuracy, or serving behavior.

AI assistance and accountability

OpenAI Codex assisted with drafting and validating this change. The human submitter must review every changed line and understand and defend the contribution end-to-end.

Summary by CodeRabbit

• Documentation
  • Added guidance for writing and reviewing Triton kernels, covering implementation practices, launch configuration, indexing, synchronization, data types, boundary validation, benchmarking, and compilation overhead.
  • Added recommendations for avoiding unintended runtime compilation and accounting for compilation and autotuning costs during performance evaluation.
  • Added agent guidance that helps apply these recommendations when implementing or reviewing Triton kernels, with clear metadata and default instructions.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review for a one-time review, or @claude review always to subscribe this PR to a review on every future push.

_{Tip: disable this comment in your organization's Code Review settings.}

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: ef26d8f3-39f1-46b1-a05d-0e00d2e021aa

📥 Commits

Reviewing files that changed from the base of the PR and between 4813ef4 and 754d69a.

📒 Files selected for processing (1)

• .agents/skills/triton-kernel-writing/SKILL.md

🚧 Files skipped from review as they are similar to previous changes (1)

• .agents/skills/triton-kernel-writing/SKILL.md

Included review availability: Your plan provides up to 10 included reviews per hour; 6 remain after this review.

---

📝 Walkthrough

Walkthrough

This change adds a triton-kernel-writing skill for vLLM. It documents Triton implementation, launch, indexing, and validation practices. An OpenAI agent configuration exposes the skill with display metadata and a default prompt.

Changes

Triton kernel skill

Layer / File(s)	Summary
Skill definition and agent wiring .agents/skills/triton-kernel-writing/SKILL.md, .agents/skills/triton-kernel-writing/agents/openai.yaml	The skill defines Triton kernel-writing guidance. The agent configuration provides display metadata and a default prompt.
Kernel implementation guidance .agents/skills/triton-kernel-writing/SKILL.md	The implementation guidance covers Triton semantics, generated-code inspection, compile-time controls, specialization behavior, and tl.debug_barrier() for write-then-read patterns.
Launch, indexing, and validation guidance .agents/skills/triton-kernel-writing/SKILL.md	The skill documents grid limits, int64 offset arithmetic, decode and prefill mapping, boundary checks, dtype selection, benchmark sweeps, and compilation overhead reporting.

Estimated code review effort: 1 (Trivial) | ~3 minutes

Merge Risk: ⚪ Minimal · up to 754d6

This change adds repository-local Triton kernel-writing guidance without changing runtime or serving behavior, and no actionable merge-blocking risk remains.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely describes the main change: adding a Triton kernel-writing skill for agents.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0…
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

Full details: Docstring Coverage

Explanation

No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check. Docstring coverage is scoped to functions touched by this diff. Analyzed 0 functions across 0 files. (1 skipped: 1 unsupported.)

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch woosuk/kernel-skill

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[WoosukKwon]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86951 for commit 754d69a2a391.