optimizing-attention-flash

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.

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

71%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is highly actionable with copy-paste code and well-sequenced workflows, but is verbose with repeated benchmark code and ships two broken reference links. Tightening redundant examples and creating the missing reference files would lift the lower dimensions.

Suggestions

Create the missing references/algorithm.md and references/advanced-features.md (or remove the links) so all navigated references resolve.

Dedupe the benchmarking/profiling code — present it once and reference it from each workflow instead of repeating near-identical timer blocks in Workflow 1 Step 3 and Workflow 2 Step 4.

Add inline feedback loops (e.g. "If speedup <2x, verify seq length ≥512 and CUDA capability ≥7.5") within the workflows rather than only in the separate Common issues section.

Dimension	Reasoning	Score
Conciseness	The ~360-line body is mostly executable code without concept re-explanation, but benchmark/profiling code is repeated across Workflow 1 (Step 3) and Workflow 2 (Step 4), and three full workflows could be tightened — matching the "mostly efficient but could be tightened" anchor.	3 / 5
Actionability	Fully executable, copy-paste-ready code throughout — SDPA quick start, flash_attn_func calls, FP8 conversion, and benchmark snippets — with concrete commands covering the common cases.	5 / 5
Workflow Clarity	Each workflow has a 4-item checklist with verification steps ("Verify speedup with profiling", "Test accuracy matches baseline"), but error-recovery feedback loops live in a separate "Common issues" section rather than inline within the workflows.	4 / 5
Progressive Disclosure	Sections are clearly signaled in an "Advanced topics" block with one-level-deep links, but 2 of 4 referenced files (algorithm.md, advanced-features.md) do not exist in the bundle, breaking navigation, and substantial inlined content could be split out.	3 / 5
	Total	15 / 20 Passed

Description

92%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is well-constructed: it states concrete capabilities, gives explicit "Use when..." triggers with natural phrasing, and carves out a distinct niche. Trigger-term coverage is strong but could add common synonyms like "OOM".

Dimension	Reasoning	Score
Specificity	Lists multiple concrete capabilities — "Optimizes transformer attention with Flash Attention", "2-4x speedup and 10-20x memory reduction", and "Supports PyTorch native SDPA, flash-attn library, H100 FP8, and sliding window attention" — giving comprehensive coverage of the capability surface.	5 / 5
Completeness	Clearly answers what ("Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction") and when ("Use when training/running transformers with long sequences... encountering GPU memory issues... or need faster inference") with concrete trigger phrases.	5 / 5
Trigger Term Quality	"Use when training/running transformers with long sequences (>512 tokens), encountering GPU memory issues with attention, or need faster inference" gives good natural keyword coverage, but misses common synonyms like "OOM" or "out of memory".	4 / 5
Distinctiveness Conflict Risk	A clear niche (Flash Attention optimization) with distinct triggers (long sequences, GPU memory issues with attention, FP8/sliding window) that minimize conflict with general ML-training skills.	5 / 5
	Total	19 / 20 Passed

Validation

81%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 13 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning
relative_links	Relative link issues: 2 missing	Warning
referenced_paths_exist	Referenced path issues: 4 missing	Warning

	Total	13 / 16 Passed

Repository: synthetic-sciences/openscience
Path: backend/cli/skills/ml-training/flash-attention/SKILL.md
Commit: 3a6c3a9

Reviewed: 1 day ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.