Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add Dogacel/auto-gpu-kernel --skill optimizegit clone --depth 1 https://github.com/Dogacel/auto-gpu-kernelWrote 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/dogacel/auto-gpu-kernel/optimize)<a href="https://agentmods.dev/skills/dogacel/auto-gpu-kernel/optimize"><img src="https://agentmods.dev/badge/skills/dogacel/auto-gpu-kernel/optimize/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/dogacel/auto-gpu-kernel/optimize"><img src="https://agentmods.dev/badge/skills/dogacel/auto-gpu-kernel/optimize.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 | $0.00027 | $0.01418 |
| Opus 5.5 | $0.00011 | $0.00567 |
| Sonnet 5.5 | $0.00005 | $0.00284 |
| Haiku 4.5 | $0.00003 | $0.00142 |
Grade A, and why
optimize 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 1mo 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 — 81 lines — stays where its author put it; the contents beside it link to each section on GitHub.
optimize — autonomous optimization loop
Iteratively improve the kernel named by config.toml. Rules in AGENTS.md are non-negotiable.
Loop
IMPORTANT: Make sure the research agent is called every 5-10 experiments to ensure we are not going in circles.
-
Assess. Read the kernel, the baseline,
experiments/summary.md,experiments/LESSONS.md. For directly relevant prior attempts, readexperiments/exp_N/result.md. If the highest-numbered folder hasplan.mdbut noresult.md, implement that plan — it's reserved (see §Folder reservation). -
Plan one change. Follow the progression ladder in
AGENTS.md§Language. Don't skip structural wins for micro-tuning. Scansummary.mdfor similar past attempts; if close, articulate what's different this time.Study workload characteristics — the distribution of every variable axis listed in
AGENTS.md§This kernel, plus any exploitable structure — and branch the kernel when a regime admits a cheaper path. Input-characteristic specialization can beat a one-size-fits-all kernel by orders of magnitude. Fair game as long as the win is real and not workload gaming.Minimize launches and copies. Fold prologue/epilogue ops (initialization, padding, masking, remapping) into the main kernel rather than calling them as separate launches. Avoid
.contiguous()when you can plumb strides into the kernel instead — each copy is both a launch and a memory round-trip. -
Implement. One optimization. Preserve DPS — the outputs are pre-allocated and passed in; never allocate them inside the kernel.
-
Validate.
kbench bench --quick(2 workloads: smallest + largest — catches shape-assumption bugs). Fix compile/correctness before proceeding. -
Measure.
kbench bench --stride 2. Before trusting the number:- Reference-latency sanity: if the ref latency is >30% off the moving median from recent
summary.mdrows, the VM is anomalous — re-run once. - Sub-5% deltas are noise on cross-VM comparison. Confirm with
kbench ab --a experiments/exp_<prev-best>/solution_fused.py. - Report latency split into small/large groups when both are present; aggregated means hide regime-specific regressions.
- If results are looking good or you are not certain due to noise, proceed with a full
kbench bench.
- Reference-latency sanity: if the ref latency is >30% off the moving median from recent
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.
- 1mo ago First seen · 81 lines · 27 tokens per session scan A 888af5b42e00
optimize is a skill published in the GitHub repository Dogacel/auto-gpu-kernel (317 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 27 tokens to every session and 1,418 once invoked, about $0.0001 per session on Opus 5.5. 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 2026-09-13.
Other skills, from other repositories
cuda-kernel-optimizer
Use when optimizing or diagnosing GPU workload performance with CUDA, CUTLASS, Triton, PyTorch, vLLM or TensorRT-LLM; interpreting NCU, Nsys or PyTorch Profiler evidence; or preparing the workload and correctness checks for such optimization. General coding, skill maintenance and conceptual GPU questions do not…
cli-anything-renderdoc
CLI harness for RenderDoc graphics debugger capture analysis.
cli-anything-nsight-graphics
Windows-first CLI harness for Nsight Graphics capture, GPU Trace summary, and ngfx-replay analysis.
colab-finetuning
Fine-tune LLMs on Google Colab GPUs directly from openscience. Connects to Colab runtimes via WebSocket bridge for remote training with Unsloth. Supports SFT, GRPO, DPO, vision, and TTS workflows on free T4 to Pro A100 GPUs.
modal-serverless-gpu
Run approved CPU or GPU work through OpenScience computejob on the user's configured Modal account. Use for isolated scientific scripts, dependency provisioning, durable outputs, logs, status, cancellation, and recovery. Never invoke the Modal SDK or CLI directly.
dstack-prototyping
Use with the dstack skill for model-serving work when the image, serving command, resources, backend/fleet choice, or service behavior is not proven. Guides task-first prototyping on real hardware, choosing fleets/backends that can reuse idle instances and caches, checking vLLM/SGLang sources, and verifying the final…