338 skills · sorted by relevance
Integrate Flash Attention into transformer models to reduce processing time and GPU memory use for long sequences. Covers PyTorch integration, advanced attention options, and benchmarks that compare speed, memory use, and output accuracy.
Optimize transformer attention, the calculation that relates tokens to one another, using Flash Attention. The skill guides PyTorch integration, GPU memory reduction, performance measurement, and output comparisons for training or inference with long sequences.
attention notification watcher（承認待ち / 判断待ち / CI 失敗等を OS 通知 + 音で能動通知）を dispatcher ペインの右側に split で常駐起動する。`.state/attention.json` が未配置なら ja 既定テンプレートを `tools/templates/attention.example.json` から自動コピーする。 起動後のペイン id は `.state/attention_pane.json` に記録し、`/org-attention-stop` から参照する。 「attention 起動」「通知監視を始めて」「watcher を立てて」等で発動。 `/org-start` からの auto-start はしない（明示起動推奨ポリシー）。
The English summary is being prepared. Showing the original description.
`/org-attention-start` で起動した attention watcher ペインを停止する。 `.state/attention_pane.json` に記録された pane_id を `list_panes` の name/role で identity 確認し、いまも attention watcher を指している場合だけ `close_pane`（pane 破棄）で 破棄する（pane_id が別ペインへ再割当て済みなら close せず stale sidecar として削除）。 「attention 止めて」「通知監視を停止」「watcher 落として」等で発動。
The English summary is being prepared. Showing the original description.
Integrate Flash Attention into Transformer models to speed up attention, the calculation that relates input elements, while reducing GPU memory use. Covers PyTorch integration, performance measurement, output comparison, and a separate H100 optimization path.
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.
The English summary is being prepared. Showing the original description.
The runbook a triage agent follows on each run over the super-admin attention dock, through the attention MCP tool. Use when a scheduled attention-triage routine wakes, when asked to triage, clean up or shrink the attention dock or admin alerts, or when the dock's card is over budget.
The English summary is being prepared. Showing the original description.
Learn how instructions, documents, tool results, and conversation history form an AI model’s context. Understand attention limits, why information in the middle can be missed, and how relevance and placement inform decisions about what to include.
Install Triton + SageAttention to accelerate ComfyUI (the sageattn attention_mode and inductor torch.compile used by WanVideoWrapper / many video graphs). Windows-first (triton-windows + woct0rdho prebuilt SageAttention wheels matched to torch/CUDA/python into the RIGHT python), plus Linux (official triton + build) and Mac (N/A → sdpa/MPS). Also covers the SAFE sdpa / no-compile fallback so an example that assumes sageattn + torch.compile still runs when these aren't installed (video-extend TRAP 5). Use when a loader crashes with "No module named 'sageattention'" or reports triton unavailable, when asked to speed up Wan/video workflows, or when deciding whether to install acceleration vs. fall back.
The English summary is being prepared. Showing the original description.
Pick the right ComfyUI startup flags for VRAM, attention, caching, and speed. The full decision matrix for OOM (--novram / --cache-none / --disable-smart-memory), shared-VRAM creep on Windows (--reserve-vram N), model-switching with big text encoders (--cache-none), high-VRAM throughput (--gpu-only / --highvram), and attention-backend selection (--use-sage-attention for speed, --use-pytorch-cross-attention as the highest-quality / Z-Image-safe fallback). Also the acceleration-stack + Blackwell/RTX 5000 (sm_120) notes. Use when a graph OOMs (especially long video like LTX 2 / WAN), when the GPU spills into shared VRAM and slows to a crawl, when switching between models eats all RAM, when Z-Image produces black/garbled output under Sage, or when deciding which attention backend to launch with. Flag names verified against upstream comfy/cli_args.py; see Sources.
The English summary is being prepared. Showing the original description.
Otimiza atenção em transformers com Flash Attention para ganho de 2-4x em velocidade e redução de 10-20x em memória. Use ao treinar/executar transformers com sequências longas (>512 tokens), ao encontrar problemas de memória GPU com atenção, ou quando precisa de inferência mais rápida. Suporta SDPA nativo do PyTorch, biblioteca flash-attn, H100 FP8 e sliding window attention.
The English summary is being prepared. Showing the original description.
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.
The English summary is being prepared. Showing the original description.
Measure and optimize claude-flow v3 startup, search, attention processing, and memory use. Compare benchmarks against performance targets, track changes over time, and identify regressions to guide further optimization.
Measure and optimize claude-flow v3 startup, search, attention processing, and memory use. Compare benchmarks against performance targets, track changes over time, and identify regressions to guide further optimization.
Investigate how transformer language models produce predictions with TransformerLens. This skill guides inspection of intermediate values and attention patterns, plus experiments that replace internal values to test their effects and identify components involved in model behavior.
Explore how coordinating agents can pass relevant working state to workers through KV cache, a model's internal attention memory. Covers task-specific selection, runtime requirements, and comparisons of accuracy, token use, and processing overhead.
Helps extend how much text a language model can process at once. Covers choosing RoPE, YaRN, ALiBi, or position interpolation, implementing position information, fine-tuning on longer documents, and managing memory during inference.
Investigate how transformer language models produce predictions with TransformerLens. This skill guides inspection of intermediate values and attention patterns, plus experiments that replace internal values to test their effects and identify components involved in model behavior.
Shape output for a reader with ADHD, then write each sentence in controlled English: action first, one word one meaning, active voice, simple tenses, state restated every turn. Invoke with /attention-control; stays on until "stop attention control".
The English summary is being prepared. Showing the original description.
Simulate readers encountering a draft for the first time to see whether they open it, where their attention fades, and what they remember. Receive reader comments alongside the text without rewrites or suggested fixes.
Calculate how to divide serving memory between recurrent state and KV cache, two memory pools used by hybrid models. Use startup logs and workload settings to recommend a ratio, estimate concurrency, and identify the limiting pool.
Set up LLM serving with structured responses, including JSON and outputs restricted to specified patterns. Covers reusing shared prompt calculations, tool calling, and multi-turn conversations through SGLang, with examples for server setup and application integration.
Build LLM serving workflows that produce JSON or other required formats. The skill covers reusing computation for shared prompts, tool calls, and ongoing conversations, with examples for server setup and structured output.
Implement paper-defined attention variants by extracting mechanism invariants, mapping tensor shapes, preserving module interfaces, validating numerical behavior, and integrating the custom attention into an existing transformer stack.
The English summary is being prepared. Showing the original description.