Official agent skill

Ir Override Ablation

by facebookexperimental in facebookexperimental/triton

Design and run Triton TTGIR debugging ablations using iroverride.

OfficialMITAuto-check passedDevelopment

Install Ir Override Ablation

skills CLI
$ npx skills add facebookexperimental/triton --skill ir-override-ablation -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install facebookexperimental/triton ir-override-ablation --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/facebookexperimental/triton.git skills-src && mkdir -p .claude/skills && cp -r skills-src/.claude/skills/ir-override-ablation .claude/skills/ir-override-ablation && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
ir-override-ablation
GitHub stars
201
Token cost
~978 tokens
SKILL.md length
471 words
Files
3 (incl. references)
Skills in repo
18
Repo updated
First seen
Licence
MIT

At a glance

Design and run Triton TTGIR debugging ablations using iroverride.

  • Works in 3 steps: Identify the target signal before editing → Get a baseline TTGIR → Decide the override mechanism
  • Reducing a provided
  • SKILL.md covers Setup, Ablation Loop, Harness Rules and Reporting, plus 1 more section
  • Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md

What it does

Ir Override Ablation is an agent skill from facebookexperimental/triton, published by the product's own GitHub organization. Design and run Triton TTGIR debugging ablations using iroverride. Use when reducing a provided or dumped TTGIR, trying user-provided or agent-generated ablation/oblation ideas, updating a test harness around iroverride, or preserving a compile/runtime failure while simplifying IR to expose a fundamental compiler or lowering gap.

Its SKILL.md is about 980 tokens, which your agent loads only when the skill is triggered. The skill folder holds 3 other files, including reference files (for example `references/harness-template.md` and `references/variant-manifest-template.md`).

It sits in Development. The repository describes itself as: Github mirror of trition-lang/triton repo. The licence is MIT.

When your agent uses it

  • Reducing a provided
  • Trying user-provided
  • Agent-generated ablation/oblation ideas
  • Updating a test harness around iroverride

Example prompts

  • “/ir-override-ablation”

Requirements

  • Python 3

Workflow steps

3 steps, taken from the first numbered list in SKILL.md.

  1. Identify the target signal before editing
  2. Get a baseline TTGIR
  3. Decide the override mechanism

What it can do on your machine

Read from SKILL.md and the folder at commit 37301d4. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    No scripts in the folder and no shell commands in SKILL.md.

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Ir Override Ablation loads about 978 tokens when it runs, and up to ~1.8k if it reads all its reference files. Until then it costs about 88 tokens; SKILL.md has 471 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~88
When it runs · the whole SKILL.md, loaded when a task matches
~978
With references · SKILL.md plus every file in references/, read only if the agent opens them
~1.8k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from facebookexperimental/triton at commit 37301d4, republished under its MIT licence (© facebookexperimental). 471 words, ~978 tokens.

Download SKILL.mdSave it as .claude/skills/ir-override-ablation/SKILL.md (or your agent's skills folder). This skill also uses 2 other files; get the full folder from GitHub.
name
ir-override-ablation
description
Design and run Triton TTGIR debugging ablations using ir_override. Use when reducing a provided or dumped TTGIR, trying user-provided or agent-generated ablation/oblation ideas, updating a test harness around ir_override, or preserving a compile/runtime failure while simplifying IR to expose a fundamental compiler or lowering gap.

IR Override Ablation

Use this skill to turn a noisy TTGIR reproducer into a smaller, evidence-backed set of IR variants. The goal is not only to simplify the IR; it is to preserve the debugging signal and keep the test oracle honest when edits intentionally change kernel semantics.

Setup

  1. Identify the target signal before editing:
    • compile failure, pass assertion, crash, hang, wrong result, or diagnostic
    • the exact command and environment that reproduces it
    • the minimum architecture requirement, for example sm90 or sm100
  2. Get a baseline TTGIR:
    • Use the provided .ttgir directly when available.
    • Otherwise dump it with:
      bash
      TRITON_ALWAYS_COMPILE=1 TRITON_KERNEL_DUMP=1 TRITON_DUMP_DIR=<dump_dir> <run_command>
  3. Decide the override mechanism:
    • Prefer triton.Config(..., ir_override="<variant.ttgir>") when the harness already uses configs/autotune or can be cheaply adapted.
    • Use TRITON_KERNEL_OVERRIDE=1 TRITON_OVERRIDE_DIR=<dump_dir> when testing against the dumped override directory layout is simpler.

Ablation Loop

  1. Build or accept a list of ideas. Each idea should state the hypothesis and the expected effect on the target signal.
  2. Copy the baseline TTGIR to a named variant file before editing.
  3. Classify the test oracle before running the variant:
    • preserve: the original numeric/reference output must still match.
    • replace: the edit intentionally changes output values; update the expected output to match the new semantics.
    • relax: success is reproducing a compile failure, crash, hang, or diagnostic; numeric equality is not the primary oracle.
    • observe: temporary exploration only; do not treat this as final evidence.
  4. Update the harness when the oracle is not preserve.
    • Example: if a store operand is replaced by a constant tensor, expected output must become that constant over the same mask/shape, not the original reference computation.
    • If an edit removes a store, expected output may need to assert unchanged sentinel values instead of comparing to the old result.
  5. Run exactly the command recorded for the variant with TRITON_ALWAYS_COMPILE=1 to avoid cache reuse.
  6. Record whether the target signal survives, weakens, disappears, or changes into a different failure.
  7. Prefer single-concept edits. When multiple edits are needed, stage them as separate variants so the final reproducer has a defensible chain of evidence.
Show full SKILL.md (125 more words)Show less

Harness Rules

  • Keep baseline and variant runs side by side until the final reproducer is chosen.
  • Make oracle-changing behavior explicit in code or comments near the assertion.
  • For wrong-result bugs, keep at least one shape that fails and one nearby shape that passes when practical.
  • For hangs, set a short external timeout and run third_party/tlx/killgpu.sh if the GPU process survives longer than expected.
  • Do not run performance benchmarks unless the user explicitly asks.

Reporting

Return a compact result matrix:

text
variant | hypothesis | oracle | result | target signal | kept?

Then state the smallest surviving reproducer, the concepts still present in the IR, and the concepts successfully removed.

References

  • Use references/variant-manifest-template.md when tracking more than a few ideas or when handing work to another agent.
  • Use references/harness-template.md when creating or adapting the Python harness around ir_override.

© facebookexperimental, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 2 other files (references) in .claude/skills/ir-override-ablation of facebookexperimental/triton.

  • SKILL.md
  • references/harness-template.md
  • references/variant-manifest-template.md

Open the folder on GitHubat commit 37301d4

Compare with similar skills

Ir Override Ablation next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Ir Override Ablation compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Ir Override Ablation this skillfacebookexperimental/triton201—~978Automated safety check: PassMIT
Vercel Composition Patternssupabase/supabase111k58 repos~726Automated safety check: PassMIT
Finishing a Development Branchobra/superpowers297k5 repos~1.9kAutomated safety check: PassMIT
Typescript Advanced Typesrolling-scopes/rsschool-app10k25 repos~4.2kAutomated safety check: PassMPL-2.0
PR Babysitteropeninterpreter/openinterpreter69k3 repos~4.2kAutomated safety check: PassApache-2.0
Code Review ChecklistshareAI-lab/learn-claude-code78k4 repos~1.1kAutomated safety check: PassMIT

Similar skills

  • Official

    React composition patterns that scale. An agent skill from supabase/supabase.

    111k GitHub starsUsed in 58 repos~726 tokens
    DevelopmentAuto-check passed
  • Walks the last step of a branch: confirm tests pass, detect the git environment, ask how to integrate, carry out your choice and clean up the worktree.

    297k GitHub starsUsed in 5 repos~1.9k tokens
    DevelopmentAuto-check passed
  • Typescript Advanced Types

    rolling-scopes/rsschool-app

    Master TypeScript's advanced type system including generics, conditional types, mapped types, template literals, and utility types for building type-safe applications.

    10k GitHub starsUsed in 25 repos~4.2k tokens
    DevelopmentAuto-check passed
  • PR Babysitter

    openinterpreter/openinterpreter

    Watches an open GitHub pull request until it merges, handling review comments, diagnosing CI failures and retrying flaky checks along the way.

    69k GitHub starsUsed in 3 repos~4.2k tokens
    DevelopmentAuto-check passed
  • Code Review Checklist

    shareAI-lab/learn-claude-code

    Reviews code against a five-part checklist covering security, correctness, performance, maintainability and testing, and reports findings in a fixed format.

    78k GitHub starsUsed in 4 repos~1.1k tokens
    DevelopmentAuto-check passed
  • Greploop

    onyx-dot-app/onyx

    Iteratively improves a PR (GitHub), MR (GitLab), or shelved changelist (Perforce) until Greptile gives it a 5/5 confidence score with zero unresolved comments.

    32k GitHub starsUsed in 4 repos~3.3k tokens
    DevelopmentAuto-check passed

More from facebookexperimental/triton

All 18 skills in this repo
  • Amd Att Trace

    facebookexperimental/triton

    Official

    Collect, validate, package, and inspect rocprofv3 Advanced Thread Trace bundles for AMD GPU kernels.

    201 GitHub stars~733 tokensUpdated today
    Auto-check passed
  • Tlx Kernel Optimization Agent

    facebookexperimental/triton

    Official

    Execute the TLX Kernel Optimization Agent CLI on a Triton or TLX kernel.

    201 GitHub stars~3.5k tokensUpdated today
    Auto-check passed
  • Compute Sanitizer

    facebookexperimental/triton

    Official

    Run NVIDIA compute-sanitizer (memcheck, racecheck, initcheck, synccheck) against a Triton/TLX kernel to find runtime memory and synchronization bugs.

    201 GitHub stars~1.6k tokensUpdated today
    Auto-check passed
  • Debug Failing GPU

    facebookexperimental/triton

    Official

    Recover from GPU-busy / GPU-unavailable failures. An agent skill from facebookexperimental/triton.

    201 GitHub stars~709 tokensUpdated today
    Auto-check passed
  • Ir Debugging

    facebookexperimental/triton

    Official

    Debug Triton compilation by dumping IR at each stage (TTIR, TTGIR, LLVM, PTX).

    201 GitHub stars~644 tokensUpdated today
    Auto-check passed
  • Kernel Perf Testing

    facebookexperimental/triton

    Official

    Run TLX kernel performance benchmarks on Hopper, Blackwell, and AMD (gfx950/CDNA4, gfx1250) GPUs.

    201 GitHub stars~1.1k tokensUpdated today
    Auto-check passed

Categories

Questions about Ir Override Ablation

What does Ir Override Ablation do?

Design and run Triton TTGIR debugging ablations using iroverride. Ir Override Ablation is an agent skill from facebookexperimental/triton, published by the product's own GitHub organization. Design and run Triton TTGIR debugging ablations using iroverride.

When should I use Ir Override Ablation?

Ir Override Ablation fits situations like: reducing a provided; trying user-provided; agent-generated ablation/oblation ideas; updating a test harness around iroverride.

How do I install Ir Override Ablation in Claude Code?

Run `npx skills add facebookexperimental/triton --skill ir-override-ablation -a claude-code`. Or copy the skill folder (.claude/skills/ir-override-ablation in facebookexperimental/triton) into .claude/skills/ir-override-ablation in your project. Claude Code loads it when a task matches its description.

How do I install Ir Override Ablation in Codex?

Run `npx skills add facebookexperimental/triton --skill ir-override-ablation -a codex`. Or copy the skill folder (.claude/skills/ir-override-ablation in facebookexperimental/triton) into .agents/skills/ir-override-ablation in your project. Codex loads it when a task matches its description.

Can I use Ir Override Ablation in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add facebookexperimental/triton --skill ir-override-ablation -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/ir-override-ablation, .gemini/skills/ir-override-ablation, .github/skills/ir-override-ablation and .opencode/skills/ir-override-ablation in your project.

What does Ir Override Ablation need to run?

SKILL.md names no scripts, command-line tools or credentials: Ir Override Ablation is instructions for the agent only. Our summary lists: Python 3.

Does Ir Override Ablation access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Ir Override Ablation safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Ir Override Ablation use?

Ir Override Ablation is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Ir Override Ablation use?

About 978 tokens (SKILL.md is roughly 3.9k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 806 tokens, read only when the agent opens those files.

What are the alternatives to Ir Override Ablation?

Skills that share tags, products or a category with Ir Override Ablation: Vercel Composition Patterns (supabase/supabase, 111k stars), Finishing a Development Branch (obra/superpowers, 297k stars), Typescript Advanced Types (rolling-scopes/rsschool-app, 10k stars) and PR Babysitter (openinterpreter/openinterpreter, 69k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Ir Override Ablation?

facebookexperimental (a GitHub organization, an official publisher) maintains it in facebookexperimental/triton, which has 201 GitHub stars. The repository holds 18 skills in this directory. The repository was last updated on October 11, 2026.

Source: facebookexperimental/triton on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.