agentmods.dev

h100-triton-rmsnorm

h100-triton-rmsnorm is a skill for Claude Code, Codex from patrick-toulme/harnessgym. It costs 53 tokens per session (2,309 once invoked), scanned A, original, Apache-2.0.

A workflow for optimizing a fused RMSNorm and SiLU gate kernel on an H100 GPU using Triton, a language for writing GPU kernels. It includes benchmark runs, configuration sweeps, rollback checks, and final correctness verification.

In plain words
What is it for?
Use it to inspect prior benchmark results, check remote GPU access, run development or final evaluations, sweep configurations, and verify or restore kernel changes.
Why use it?
It helps compare optimization attempts safely while preserving correctness and recovering from candidates that perform worse.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one.

Needs its repository: it runs a file that does not travel with it, so clone the repository first. The line is python3 .harnessgym/mcp/h100_triton_rmsnorm/server.py --self-test.

not rated 41repo 2mo ago A scan 53 tokens original Apache-2.0

Good fit Use it to inspect prior benchmark results, check remote GPU access, run development or final evaluations, sweep configurations, and verify or restore kernel changes.

Compare 6 skills from other repositories ↓
View source ↗ patrick-toulme/harnessgym harnessgym.com
Install

Getting it into your agent

It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.

Clone the repo
git clone --depth 1 https://github.com/patrick-toulme/harnessgym
agentmods
npx agentmods add skills/patrick-toulme/harnessgym/h100-triton-rmsnorm

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for h100-triton-rmsnorm

README.md
[![agentmods](https://agentmods.dev/badge/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm/github.svg)](https://agentmods.dev/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm)
Your own site
<a href="https://agentmods.dev/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm"><img src="https://agentmods.dev/badge/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for h100-triton-rmsnorm

Your own site · 80×15
<a href="https://agentmods.dev/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm"><img src="https://agentmods.dev/badge/skills/patrick-toulme/harnessgym/h100-triton-rmsnorm.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 53 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 2,309 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00053 $0.02309
Opus 5 $0.00026 $0.01154
Sonnet 5 $0.00011 $0.00462
Haiku 4.5 $0.00005 $0.00231

Measured 12d ago against content hash ed0f57417b05, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-12, from the pricing page.

Security

Grade A, and why

h100-triton-rmsnorm scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

examples/triton_rmsnorm_h100_harness_artifacts/.harnessgym/skills/h100-triton-rmsnorm/SKILL.md · 117 lines

How it starts

The opening of the file, as written. The whole thing — 117 lines — stays where its author put it; the contents beside it link to each section on GitHub.

H100 Triton RMSNorm Harness

Use this skill at the start of a fresh attempt on this workspace.

First Actions

  1. Inspect active context and prior results with the MCP tool inspect_context.
  2. If HARNESSGYM_GPU_HOST is set or local CUDA is uncertain, run remote_health_check before any objective sweep.
    • It checks SSH reachability, remote scratch space, and nvidia-smi visibility without starting a tar sync.
    • If it reports failure_stage="ssh", stop and record the infrastructure blocker instead of spending time on benchmark commands that will only surface tar: Write error.
  3. Use run_objective for fast iteration:
    • mode="dev", verifier=false runs benchmark.py --json --mode dev.
    • Objective metric is best_us; lower is better.
  4. Use run_objective for authoritative comparison:
    • mode="final", verifier=true runs verifier.py --json --mode final.
    • Preserve correctness: every case must pass max_abs <= 7.5e-3.
  5. Before ending on a risky candidate, call guarded_final_verify with restore_on_regression=true.
    • It compares against the best checkpoint score and restores kernel.py / kernel_config.json if the candidate regresses.
  6. Record candidate runs in history and compare with rank_history.
  7. For combined source and launch search, prefer joint_source_launch_search before manual patching.
    • Start with a small dev-mode filter, then confirm top candidates with final verifier mode.
    • It can search current, exact SiLU (exp, sigmoid, exp2), and rational SiLU variants across combined 2048/4096 launch overlays while restoring files by default.
  8. For source-level math variants, prefer sweep_silu_variants over manual patching.
    • It can test exp, sigmoid, and exp2 SiLU forms while preserving/restoring kernel.py.
    • Cross it with a small list of launch overlays when investigating the 8192 case.
  9. For approximate SiLU math, do not edit kernel.py by hand first.
    • Run probe_silu_approximations with mode="all" to check toy plus dev/final-shape proxy tolerance.
    • Then run sweep_silu_approximations; it is rollback-safe and skips variants that fail the numerical prefilter.
  10. Before trusting a borderline winner, run repeat_objective in final verifier mode.
  • Use median_score and per-case spreads to avoid keeping one lucky timing sample.

Read the full file on GitHub · 117 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 12d ago First seen · 117 lines · 53 tokens per session scan A ed0f57417b05

Subscribe to this mod's changes

h100-triton-rmsnorm is a skill published in the GitHub repository patrick-toulme/harnessgym (41 stars, last pushed 2mo ago), licensed Apache-2.0. It adds 53 tokens to every session and 2,309 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Claude CodeLanguagesScientificagentsai-agentsbenchmarkingcodexcoding-agents
Related

Other skills, from other repositories

cuopt-numerical-optimization-api

LP, MILP, and QP (beta) with cuOpt — Python, C, and CLI. Use when the user is solving LP, MILP, or QP with any cuOpt interface.

NVIDIA/skills · 51 tokens

nanoresearch-experiment

Generate a Python code skeleton from an experiment blueprint.

OpenRaiser/NanoResearch · 15 tokens

cupynumeric-install

Install and verify cuPyNumeric for Python — requirements, commands, verification. Source builds are out of scope.

NVIDIA/skills · 27 tokens

shapely-compute

Computational geometry with Shapely - create geometries, boolean operations, measurements, predicates.

parcadei/Continuous-Claude-v3 · 24 tokens

add-rcpp-integration

Add Rcpp or RcppArmadillo integration to an R package for high-performance C++ code. Covers setup, writing C++ functions, RcppExports generation, testing compiled code, and debugging. Use when an R function is too slow and profiling confirms a bottleneck, when you need to interface with existing C/C++ libraries, or…

pjt222/agent-almanac · 93 tokens

pycse

Skill "pycse" from ItamarZand88/awesome-agent-conventions, covering pycse - python computations in science and engineering, core capabilities, 1. nonlinear regression and curve fitting, 2. design of experiments (doe) and 3. uncertainty quantification with dpose.

ItamarZand88/awesome-agent-conventions · 0 tokens