Getting it into your agent
There is no command for this one: it runs only inside a plugin, and the catalogue could not identify which plugin ships it. The source is linked below.
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.
[](https://agentmods.dev/skills/syhya/mlsys26-flashinfer-contest/flashinfer-b200-contest-optimizer)<a href="https://agentmods.dev/skills/syhya/mlsys26-flashinfer-contest/flashinfer-b200-contest-optimizer"><img src="https://agentmods.dev/badge/skills/syhya/mlsys26-flashinfer-contest/flashinfer-b200-contest-optimizer/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.
<a href="https://agentmods.dev/skills/syhya/mlsys26-flashinfer-contest/flashinfer-b200-contest-optimizer"><img src="https://agentmods.dev/badge/skills/syhya/mlsys26-flashinfer-contest/flashinfer-b200-contest-optimizer.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00134 | $0.10243 |
| Opus 5 | $0.00067 | $0.05122 |
| Sonnet 5 | $0.00027 | $0.02049 |
| Haiku 4.5 | $0.00013 | $0.01024 |
Grade A, and why
flashinfer-b200-contest-optimizer 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 11d 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.
How it starts
The opening of the file, as written. The whole thing — 647 lines — stays where its author put it; the contents beside it link to each section on GitHub.
FlashInfer B200 Contest Optimizer
Overview
Use this skill to optimize MLSYS26 FlashInfer contest operators for B200, benchmark on Modal, pair each local gate with a same-round baseline run, extract compact NCU metrics from scratch artifacts, and keep reproducible artifacts for submission.
Do not start profiling or candidate writing until the reference repositories for the active operator have been refreshed and inspected. Reuse or adapt an existing operator, kernel primitive, or optimization idea when the reference scan shows one is already close to the active problem.
Assume CUDA 13.2 is the active optimization baseline. When Triton reaches a control-surface ceiling (cluster attributes, DSMEM, launch-policy control, or launch-latency overlap), move the hotspot to solution/cuda/kernel.cu instead of forcing a Triton-only path.
Treat the official contest evaluator as the final ground truth:
- Docker image:
flashinfer/flashinfer-ci-cu132:latest - hardware: bare-metal B200
- GPU clocks: locked to max with
nvidia-smi -ac 3996,1965 - timing path:
flashinfer-benchwithcupti-python - process model:
--use-isolated-runner
Deliverables per optimization round:
- A measurable latency improvement with repeat validation (latency is the primary metric — speedup is secondary due to unstable reference baseline).
- A shape-aware NCU matrix showing
band -> workload_uuid -> dominant kernel -> bottleneck -> limiter. - A Markdown benchmark summary showing
avg_latency_msandp95_latency_msas primary comparison columns, withavg_speedupas context only. - A scratch-safe archive: every full run, single-workload run, NCU report, and candidate kernel copied out of the project root before the next experiment.
- A round decision that explicitly says whether a candidate was archived only, promoted globally, or rejected with the repo restored to the previous best.
Reference-First Reconnaissance
Treat reference-repo refresh and code reconnaissance as a mandatory stage 0, not optional background reading.
What ships with it
11 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
- agents/openai.yaml 1.0 KB
- references/b200-architecture-notes.md 2.2 KB
- references/contest-evaluation-environment.md 3.7 KB
- references/cuda-13-2-notes.md 2.6 KB
- references/cuda-book-checklist.md 1.1 KB
- references/parallel-eval-and-naming.md 2.5 KB
- references/reference-repo-recon.md 3.9 KB
- scripts/archive_modal_artifacts.py 8.3 KB runs code
- scripts/extract_ncu_metrics.py 13 KB runs code
- scripts/record_best_result.py 12 KB runs code
- scripts/run_modal_candidates_parallel.py 37 KB runs code
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.
- 11d ago First seen · 647 lines · 134 tokens per session scan A 9f4b0795b595
flashinfer-b200-contest-optimizer is a skill published in the GitHub repository syhya/mlsys26-flashinfer-contest (22 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 134 tokens to every session and 10,243 once invoked, about $0.0007 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.
Other skills, from other repositories
paper2poster
Convert academic papers (PDF) into conference posters (HTML/PNG). You are the conductor: you decide what each section needs — an original paper figure or text — write the outline, hand-author the poster HTML, and iterate on the render using your own visual read and a blind-reader content quiz. Use when the user wants…
paper2slides
Turn an academic paper PDF into a presentation deck (.pptx) end-to-end. Use this skill whenever the user wants to "make slides from a paper", "generate a deck from this PDF", "make a PPT from this paper", "generate slides from a PDF document", "make a deck from a research paper", or supplies a research paper PDF and…
5writing
A paper-writing workflow for mathematical modelling competitions, supporting both Typst and LaTeX. It uses analysis, results, figures, and a competition template to assemble the document.
6verity
A final-checking workflow for mathematical modelling competition papers, using either Typst or LaTeX to build the document. It checks the finished paper and its supporting files before submission.
1start-mathmodel
A starting workflow for mathematical-modeling competitions, which are contests where teams use mathematics, data, and code to solve real-world problems. It asks for preferences, creates planning files, and coordinates later analysis, modeling, coding, charting, diagram, writing, and verification stages.
4drawio
A diagram-making stage for mathematical-modeling projects. It uses existing analysis and results reports to create technical route maps, solution flowcharts, model diagrams, and data-processing diagrams in Drawio, a tool for creating structured diagrams, then exports them as PDF files.