Agent Skills: 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.
[OptimizationFlashAttentionAttentionOptimizationMemoryEfficiencySpeedOptimizationLongContextPyTorchSDPAH100FP8Transformers]
ID: PoorRican/dotfiles/optimizing-attention-flash
Install this agent skill to your local
Skill Files
The files that make up optimizing-attention-flash.
More agent skills
- azure-communication-callautomation-javaBuild call automation workflows with Azure Communication Services Call Automation Java SDK. Use when implementing IVR systems, call routing, call recording, DTMF recognition, text-to-speech, or AI-powered call flows.
- feishu-toolsUse the Feishu/Lark plugin MCP tools for Feishu documents, spreadsheets, knowledge content, and other matching workspace operations.
- githubUse gh for GitHub issues, PR status, CI/logs, comments, reviews, releases, and API queries.
- data-client-reactRead, change and render async and shared state in React with @data-client/react: any data from an API or other async source, and global state shared across components (it has to persist somewhere, which makes it async). Use it for loading and error states, mutations and forms, filtered or derived views, and debugging data that looks wrong or stale on screen. Use it in place of fetch in useEffect, or app-wide useState and context; its main hooks are useSuspense, useQuery and useController.
- trailmark-structuralRuns full Trailmark structural analysis by building a graph, running `preanalysis()`, and reporting hotspots, taint, blast radius, privilege boundaries, attack surface, and version-gated Trailmark 0.4+/0.5+ data such as proxy counts, subgraph edges, type/reference summaries, and entrypoint attributes. Use when vivisect needs detailed structural data for a target. Triggers: structural analysis, blast radius, taint analysis, complexity hotspots, proxy nodes, type references.
- self-reviewWeekly pass over the evlog repository and Evi's own surface, in two halves: what has drifted out of coherence (a capability wired but never consumed, code contradicting a written guide, a description promising a tool the allowlist lacks), and what is missing (a capability worth having, a manual step worth automating, a gap in evlog users keep hitting). Load this when the self-review schedule fires, or when Hugo asks for a self-review, an audit, ideas for what Evi should do next, or what is inconsistent in the repo.