optimizing-attention-flash

Orchestra-Research/AI-Research-SKILLs AI & ML engineering

optimizing-attention-flash is a skill in Orchestra-Research/AI-Research-SKILLs, last changed 20 Nov 2025. It can run shell commands, reach the network and name web addresses — read the findings before installing.

From the author

Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction. Use when training/running transformers with long sequences (>512 tokens), encountering GPU memory issues with attention, or need faster inference. Supports PyTorch native SDPA, flash-attn library, H100 FP8, and sliding window attention.

How much a skill can do, not whether it is risky.

Pattern check

Nothing flagged

That isn't proof the skill is harmless.

Commit 773a529 · 3 Nov 2026

How the check works

Pattern check: nothing flagged

That isn't proof the skill is harmless.

Install

Each tab says whether it installs commit 773a529, the one we checked, or the latest version.

OpenCode also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Goose also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Windsurf (Devin Desktop) also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Zed also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Roo Code also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Factory Droid also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Junie also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

Amp also looks in .agents/skills/ and ~/.agents/skills/, so use the Any agent steps.

No install command here reaches Kiro. Copy the skill's folder into .kiro/skills/ in your project, or ~/.kiro/skills/ for all your projects. Orchestra-Research/AI-Research-SKILLs/10-optimization/flash-attention (opens in a new tab)

Antigravity also looks in .agents/skills/, so the Any agent steps reach it with This project. For All projects, see where Antigravity looks.

Cline also looks in .claude/skills/, so the Claude Code steps reach it with This project. For All projects, see where Cline looks.

No tab for your agent? Use Any agent for OpenCode, Goose, Windsurf (Devin Desktop), Zed, Roo Code, Factory Droid, Junie and Amp. With This project, Any agent also reaches Antigravity and Claude Code reaches Cline. Kiro needs the folder copied by hand.

Pinned to commit 773a529, the one we checked.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .claude/skills &&mkdir .claude/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .claude/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

If Claude Code was already running and this created the skills folder, run /reload-skills.

Claude Code docs

Pinned to commit 773a529, the one we checked.

Skills in the Claude apps belong to your account rather than to a project.

Make optimizing-attention-flash.zip, with the skill folder at the top level:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&out="$(mktemp -d)" &&cp -R "$tmp"/10-optimization/flash-attention "$out"/optimizing-attention-flash &&(cd "$out" && zip -qr - optimizing-attention-flash) > optimizing-attention-flash.zip &&rm -rf "$tmp" "$out"

In Claude, open Customize > Skills, select +, then + Create skill > Upload a skill, and choose optimizing-attention-flash.zip.

On Free, Pro and Max plans, turn on Code execution and file creation (Create and edit files in newer settings) under Settings > Capabilities.

Claude apps docs

Pinned to commit 773a529, the one we checked.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .agents/skills &&mkdir .agents/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .agents/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

Codex picks up new skills on its own; restart it if this one does not appear.

Codex docs

Pinned to commit 773a529, the one we checked.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .cursor/skills &&mkdir .cursor/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .cursor/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

Cursor docs

Pinned to commit 773a529, the one we checked.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .github/skills &&mkdir .github/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .github/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

In a running Copilot CLI session, run /skills reload.

Or install it with gh skill, which also installs the commit we checked. Needs GitHub CLI 2.90 or later.

gh skill install Orchestra-Research/AI-Research-SKILLs 10-optimization/flash-attention/SKILL.md@773a52944ba4747a18bd4ae9ade53fff041adcbc --agent github-copilot --scope project

Copilot & VS Code docs

Pinned to commit 773a529, the one we checked.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .gemini/skills &&mkdir .gemini/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .gemini/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

In a running session, run /skills reload.

Or use Gemini CLI’s own installer. Installs the latest version, not the one we checked.

gemini skills install https://github.com/Orchestra-Research/AI-Research-SKILLs.git --path 10-optimization/flash-attention --scope workspace

Gemini CLI docs

Pinned to commit 773a529, the one we checked.

OpenCode, Goose, Windsurf (Devin Desktop), Zed, Roo Code, Factory Droid, Junie, Amp and Antigravity read .agents/skills in a project.

From your project's root folder:

tmp="$(mktemp -d)" &&git clone -q --depth 1 --filter=blob:none --no-checkout https://github.com/Orchestra-Research/AI-Research-SKILLs.git "$tmp" &&git -C "$tmp" fetch -q --depth 1 origin 773a52944ba4747a18bd4ae9ade53fff041adcbc &&git -C "$tmp" sparse-checkout set -- 10-optimization/flash-attention &&git -C "$tmp" checkout -q 773a52944ba4747a18bd4ae9ade53fff041adcbc &&mkdir -p .agents/skills &&mkdir .agents/skills/optimizing-attention-flash &&cp -R "$tmp"/10-optimization/flash-attention/. .agents/skills/optimizing-attention-flash/ &&rm -rf "$tmp"

Or use the skills CLI from skills.sh. Installs the latest version, not the one we checked. The skills CLI reports installs to skills.sh unless you set DISABLE_TELEMETRY=1.

npx skills@1.7.4 add Orchestra-Research/AI-Research-SKILLs --skill optimizing-attention-flash

Agent Skills docs

Read SKILL.md first

MIT

Licence

14.2k

Repo stars

Folder last changed

Publisher

Files

Checked

3 Nov 2026

What it can touch

From the files and frontmatter at commit 773a529. This is how much a skill can do, not whether it is risky.

Runs shell commands
Its instructions or files give the agent commands to run on your computer.
Reaches the network
It sends or fetches data over the internet, for example with curl, an HTTP library or a package install.
Names web addresses
It mentions outside web addresses the agent may visit. We list them but don't follow them.
Outside links we did not follow

Pattern check

Pattern check: nothing flagged · commit 773a529 · 3 Nov 2026

We checked every text file in this skill, at commit 773a529 on 3 Nov 2026, against 45 published patterns such as download-and-run commands, hidden text and credential access. We don't run the skill, we don't read the web pages it points to, and we can't judge everything its instructions might lead an agent to do. In published tests, pattern checks like this one missed most malicious skills. Read the skill before you install it — especially anything it asks you to download or run.

How the check works Report an incorrect finding

Outside links we did not follow

https://tridao.me/blog/2024/flash3/
https://github.com/Dao-AILab/flash-attention
https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html

Format check

Each agent reads SKILL.md a little differently. We check the file against each one's published rules; we don't run the skill in any agent.

Agent Skills specFails
The open format from agentskills.io, held to the strictest version of each rule among the agents we check.
- Error
  The name "optimizing-attention-flash" doesn't match its folder "flash-attention". The spec requires them to match. N4
  Fix: Set name: flash-attention, or rename the folder to optimizing-attention-flash.
- Warning
  The frontmatter uses [ ] or { } values. Most agents accept them, but the spec's validator (skills-ref) rejects the file. F4
  Fix: Write lists as "- item" lines and maps as indented key: value lines, without anchors or tags.
- Warning
  The description contains < or >. Uploads to the Claude apps and API reject it. D3
  Fix: Remove < and >, or reword (for example "less than").
- Warning
  Fields outside the spec: version, author, tags and dependencies. Uploads to the Claude apps and API, and the spec's validator, reject the file; most agents ignore these fields. O1
  Fix: Delete them for a portable skill, or keep them if you only target the agents that read them.
- Warning
  Links to files that aren't in the skill folder: line 345: references/algorithm.md and line 347: references/advanced-features.md. B3
  Fix: Add the missing files or correct the paths, relative to SKILL.md.
Claude CodePasses with warnings
Claude Code reads its own extra fields, ignores unknown ones, and still loads a skill whose frontmatter is broken, without its fields.
- Warning
  Links to files that aren't in the skill folder: line 345: references/algorithm.md and line 347: references/advanced-features.md. B3
  Fix: Add the missing files or correct the paths, relative to SKILL.md.
- Note
  The name "optimizing-attention-flash" doesn't match its folder "flash-attention". Claude Code shows the skill as /optimizing-attention-flash, and the folder name also invokes it. N4
  Fix: Set name: flash-attention, or rename the folder to optimizing-attention-flash.
- Note
  Claude Code ignores fields it doesn't know: version, author, tags and dependencies. O1
  Fix: Check the spelling against Claude Code's field list, or delete them.
Claude apps & API uploadFails
Uploading to claude.ai or the Claude API, or packaging with Anthropic's script, accepts only the six spec fields and stops on anything else.
- Error
  The name "optimizing-attention-flash" doesn't match its folder "flash-attention". claude.ai refuses an upload whose folder name doesn't match the skill name. N4
  Fix: Set name: flash-attention, or rename the folder to optimizing-attention-flash.
- Error
  The description contains < or >. Anthropic's validator rejects any angle bracket there. D3
  Fix: Remove < and >, or reword (for example "less than").
- Error
  Upload and packaging fail on fields outside the spec: version, author, tags and dependencies. O1
  Fix: Delete them, and keep your own data under metadata as text values.
- Warning
  Links to files that aren't in the skill folder: line 345: references/algorithm.md and line 347: references/advanced-features.md. B3
  Fix: Add the missing files or correct the paths, relative to SKILL.md.
- Note
  The description is 335 characters. The claude.ai Help Center says 200 at most, though Anthropic's API docs and validator allow 1,024. D4
  Fix: If the claude.ai upload refuses it, trim the description to 200 characters.
OpenAI (Codex & API)Passes with warnings
Codex finds SKILL.md by its exact name and needs a description; OpenAI's Skills API checks the frontmatter against the spec and caps uploads.
- Warning
  The name "optimizing-attention-flash" doesn't match its folder "flash-attention". OpenAI's own upload example uses a differently named folder, but the spec it follows says they must match. N4
  Fix: Set name: flash-attention, or rename the folder to optimizing-attention-flash.
- Warning
  Fields outside the spec: version, author, tags and dependencies. Codex ignores them; OpenAI doesn't say whether its upload accepts them. O1
  Fix: Delete them, or move your own data under metadata as text values.
- Warning
  Links to files that aren't in the skill folder: line 345: references/algorithm.md and line 347: references/advanced-features.md. B3
  Fix: Add the missing files or correct the paths, relative to SKILL.md.
VS Code & CopilotPasses with warnings
VS Code reads skills one folder below .github/skills, .claude/skills or .agents/skills and falls back to the folder name when the name is missing or differs; GitHub's Copilot docs require the exact name SKILL.md.
- Warning
  The name "optimizing-attention-flash" doesn't match its folder "flash-attention". VS Code uses the folder name, flash-attention, instead. N4
  Fix: Set name: flash-attention, or rename the folder to optimizing-attention-flash.
- Warning
  Links to files that aren't in the skill folder: line 345: references/algorithm.md and line 347: references/advanced-features.md. B3
  Fix: Add the missing files or correct the paths, relative to SKILL.md.
- Note
  VS Code ignores fields it doesn't support: version, author, tags and dependencies. O1
  Fix: Delete them unless another agent you target reads them.

SKILL.md

Read SKILL.md on GitHub

Files

3 files

Path	Size
SKILL.md (opens in a new tab)	10 KB
references/benchmarks.md (opens in a new tab)	7 KB
references/transformers-integration.md (opens in a new tab)	7.3 KB

More from Orchestra-Research/AI-Research-SKILLs

Skill	What it can touch	Pattern check	Licence	Updated
academic-plottingGenerates publication-quality figures for ML papers from research context. Given a paper section or description, extracts system components and relationships to generate architecture diagrams via Gemini. Given experiment results or data, auto-selects chart type and generates data-driven figures via matplotlib/seaborn. Use when creating any figure for a conference paper. Names web addresses Pattern check: Nothing flaggedMITUpdated 25 Mar 2026	Names web addresses	Pattern check: Nothing flagged	MIT	25 Mar 2026
ara-compilerCompiles any research input — PDF papers, GitHub repositories, experiment logs, code directories, or raw notes — into a complete Agent-Native Research Artifact (ARA) with cognitive layer (claims, concepts, heuristics), physical layer (configs, code stubs), exploration graph, and grounded evidence. Use when ingesting a paper or codebase into a structured, machine-executable knowledge package, building an ARA from scratch, or converting research outputs into a falsifiable, agent-traversable form. Instructions only Pattern check: Nothing flaggedMITUpdated 27 Apr 2026	Instructions only	Pattern check: Nothing flagged	MIT	27 Apr 2026
ara-research-managerRecords research provenance as a post-task epilogue, scanning conversation history at the end of a coding or research session to extract decisions, experiments, dead ends, claims, heuristics, and pivots, and writing them into the ara/ directory with user-vs-AI provenance tags. Use as a session epilogue — never during execution — to maintain a faithful, auditable trace of how a research project actually evolved. Runs shell commands Pattern check: Nothing flaggedMITUpdated 27 Apr 2026	Runs shell commands	Pattern check: Nothing flagged	MIT	27 Apr 2026
ara-rigor-reviewerPerforms ARA Seal Level 2 semantic epistemic review on Agent-Native Research Artifacts, scoring six dimensions (evidence relevance, falsifiability, scope calibration, argument coherence, exploration integrity, methodological rigor) and producing a constructive, severity-ranked report with a Strong Accept-to-Reject recommendation. Use after Level 1 structural validation passes, when an ARA needs an objective epistemic critique before publication or release. Instructions only Pattern check: Nothing flaggedMITUpdated 27 Apr 2026	Instructions only	Pattern check: Nothing flagged	MIT	27 Apr 2026
audiocraft-audio-generationPyTorch library for audio generation including text-to-music (MusicGen) and text-to-sound (AudioGen). Use when you need to generate music from text descriptions, create sound effects, or perform melody-conditioned music generation. Downloads software +3 more Pattern check: Nothing flaggedMITUpdated 25 Nov 2025	Downloads software Runs shell commands +2 more	Pattern check: Nothing flagged	MIT	25 Nov 2025
autogpt-agentsAutonomous AI agent platform for building and deploying continuous agents. Use when creating visual workflow agents, deploying persistent autonomous agents, or building complex multi-step AI automation systems. Touches credentials +3 more Pattern check: Nothing flaggedMITUpdated 25 Nov 2025	Touches credentials Runs shell commands +2 more	Pattern check: Nothing flagged	MIT	25 Nov 2025

All skills in Orchestra-Research/AI-Research-SKILLs

Other AI & ML engineering skills

Same category, most repo stars first.

Skill	What it can touch	Pattern check	Licence	Updated
agent-architecture-auditFull-stack diagnostic for agent and LLM applications. Audits the 12-layer agent stack for wrapper regression, memory pollution, tool discipline failures, hidden repair loops, and rendering corruption. Produces severity-ranked findings with code-first fixes. Essential for developers building agent applications, autonomous loops, or any LLM-powered feature. Use when an agent or LLM feature misbehaves and the failing layer is unknown, or before shipping an agent stack. affaan-m/ECC Runs shell commands Pattern check: Nothing flaggedMITUpdated 11 Aug 2026	Runs shell commands	Pattern check: Nothing flagged	MIT	11 Aug 2026
agent-evalHead-to-head comparison of coding agents (Claude Code, Aider, Codex, etc.) on custom tasks with pass rate, cost, time, and consistency metrics. Use when choosing between coding agents, or when a change to an agent setup needs measured pass rate, cost, and time rather than an impression. affaan-m/ECC Runs shell commands +1 more Pattern check: Nothing flaggedMITUpdated 11 Aug 2026	Runs shell commands Names web addresses	Pattern check: Nothing flagged	MIT	11 Aug 2026
claude-apiReference for the Claude API / Anthropic SDK — model ids, pricing, params, streaming, tool use, MCP, agents, caching, token counting, model migration. TRIGGER — read BEFORE opening the target file; don't skip because it "looks like a one-liner" — whenever: the prompt names Claude/Anthropic in any form (Claude, Anthropic, Fable, Opus, Sonnet, Haiku, `anthropic`, `@anthropic-ai`, `claude-`, `us.anthropic.`, `[1m]`); the user asks about an LLM (pricing/model choice/limits/caching) — never answer from memory; OR the task is LLM-shaped with provider unstated (agent/MCP/tool-definition/multi-agent/RAG/LLM-judge/computer-use; generate/summarize/extract/classify/rewrite/converse over NL; debugging refusals/cutoffs/streaming/tool-calls/tokens). SKIP only when another provider is being worked on (overrides all triggers): OpenAI/GPT/Gemini/Llama/Mistral/Cohere/Ollama named in the query; OR `grep -rE 'openai\|langchain_openai\|google.generativeai\|genai\|mistralai\|cohere\|ollama'` over the project hits (run this grep FIRST if no provider named — don't Read the file). anthropics/skills Touches credentials +6 more Pattern check: 1 item to reviewApache 2.0Updated 9 Oct 2026	Touches credentials Downloads software +5 more	Pattern check: 1 item to review	Apache 2.0	9 Oct 2026
mcp-builderGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK). anthropics/skills Touches credentials +4 more Pattern check: Nothing flaggedApache 2.0Updated 28 Oct 2026	Touches credentials Runs shell commands +3 more	Pattern check: Nothing flagged	Apache 2.0	28 Oct 2026
benchmark-modelsCross-model benchmark for gstack skills. (gstack) garrytan/gstack Touches credentials +2 more Pattern check: 1 item to reviewMITUpdated 30 Oct 2026	Touches credentials Runs shell commands +1 more	Pattern check: 1 item to review	MIT	30 Oct 2026
eval-planPreregister a paid eval (bars, arms, decoys, stop rules), dry-run it at zero cost, then run a three-item priced pilot whose spend table gates approval. (gstack) garrytan/gstack Runs shell commands +1 more Pattern check: 1 item to reviewMITUpdated 2 Nov 2026	Runs shell commands Pre-approves tools	Pattern check: 1 item to review	MIT	2 Nov 2026

At a glance

What it can touch

Pattern check

Install

Claude Code

Claude apps

Codex

Cursor

Copilot & VS Code

Gemini CLI

Any agent

What it can touch

Pattern check

Outside links we did not follow

Format check

SKILL.md

Files

More from Orchestra-Research/AI-Research-SKILLs

Other AI & ML engineering skills