optimize

An iterative workflow for improving a Triton GPU program that combines sparse operations. It reads previous experiments, changes one part of the implementation at a time, and measures the result under the project’s rules.

In plain words
What is it for?
Use it to optimize `solution/triton/sparse_fused.py` against its baseline, plan kernel changes, inspect experiment records, and periodically call a research agent to check that the work is still making progress.
Why use it?
It organizes repeated performance experiments so improvements are based on earlier results instead of repeated guesses. It also checks workload patterns such as sequence length and batch size when choosing an optimization.

Command for Claude Code

Install

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.

agentmods
npx agentmods add commands/dogacel/auto-gpu-kernel/optimize
Clone the repo
git clone --depth 1 https://github.com/Dogacel/auto-gpu-kernel

Made for: Claude Code.

Per session 0 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 1,448 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
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 $0.00000 $0.01448
Opus 5 $0.00000 $0.00724
Sonnet 5 $0.00000 $0.00290
Haiku 4.5 $0.00000 $0.00145

Measured 2d ago against content hash 7d57b7c9b449, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

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 2d 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.

dsa_sparse_attention_h16_ckv512_kpe64_topk2048_ps64/.claude/commands/optimize.md · 71 lines

How it starts

The opening of the file, as written. The whole thing — 71 lines — stays where its author put it; the contents beside it link to each section on GitHub.

/optimize — autonomous optimization loop

Iteratively improve solution/triton/sparse_fused.py. Rules in CLAUDE.md are non-negotiable.

Loop

IMPORTANT: Make sure research agent is called every 5-10 experiments to ensure we are not going in circles.

  1. Assess. Read sparse_fused.py, sparse_baseline.py, experiments/summary.md, experiments/LESSONS.md. For directly relevant prior attempts, read experiments/exp_N/result.md. If the highest-numbered folder has plan.md but no result.md, implement that plan — it's reserved (see §Folder reservation).

  2. Plan one change. Early progression: PyTorch → tiled Triton → fused → tile tuning → alternative tilings. Don't skip structural wins for micro-tuning. Scan summary.md for similar past attempts; if close, articulate what's different this time.

    Study workload characteristicsseq_len distribution across the 128 workloads, batch_size, max_num_pages, 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.

  3. Implement. One optimization. Preserve DPS (don't allocate output/lse inside the kernel).

  4. Validate. /benchmark quick (2 workloads: smallest + largest — catches shape-assumption bugs). Fix compile/correctness before proceeding.

  5. Measure. /benchmark stride 2. Before trusting the number:

    • Reference-latency sanity: if the ref latency is >30% off the moving median from recent summary.md rows, the VM is anomalous — re-run once.
    • Sub-5% deltas are noise on cross-VM comparison. Confirm with modal run scripts/ab_benchmark.py::run --a experiments/exp_<prev-best>/sparse_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 /benchmark full.

Read the full file on GitHub · 71 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. 2d ago First seen · 71 lines · 0 tokens per session scan A 7d57b7c9b449

Subscribe to this mod's changes

optimize is a command published in the GitHub repository Dogacel/auto-gpu-kernel (157 stars, last pushed 11d ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 1,448 tokens. 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.