Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/sandyup/comfyui-mcpnpx agentmods add skills/sandyup/comfyui-mcp/triton-sageattentionWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/sandyup/comfyui-mcp/triton-sageattention)<a href="https://agentmods.dev/skills/sandyup/comfyui-mcp/triton-sageattention"><img src="https://agentmods.dev/badge/skills/sandyup/comfyui-mcp/triton-sageattention/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/sandyup/comfyui-mcp/triton-sageattention"><img src="https://agentmods.dev/badge/skills/sandyup/comfyui-mcp/triton-sageattention.svg" alt="Reviewed on agentmods" width="80" height="20"></a>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.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00170 | $0.05610 |
| Opus 5 | $0.00085 | $0.02805 |
| Sonnet 5 | $0.00034 | $0.01122 |
| Haiku 4.5 | $0.00017 | $0.00561 |
Grade C, and why
triton-sageattention scanned grade C with 1 finding 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 10d 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.
Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
rm -rf ~/.triton This is a copy
86% identical to triton-sageattention — 270 lines differ, which has more behind it and is treated as the original. This page carries a canonical link to it rather than competing with it.
How it starts
The opening of the file, as written. The whole thing — 384 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Triton + SageAttention (ComfyUI acceleration)
See also
comfyui-launch-flagsfor the full attention / VRAM / cache flag matrix. Note the Z-Image exception: Z-Image is broken under--use-sage-attention→ launch it with--use-pytorch-cross-attentioninstead.
Overview
Two optional accelerators that many modern video graphs (especially kijai's ComfyUI-WanVideoWrapper) reference by default:
- SageAttention (
import sageattention) — a quantized attention kernel. Selected via a node'sattention_mode = sageattn(WanVideoWrapper) or ComfyUI's--use-sage-attentionstartup flag. ~20–40% faster sampling on supported NVIDIA GPUs. - Triton — the GPU kernel compiler that inductor
torch.compileneeds. WanVideoWrapper'sWanVideoTorchCompileSettings(and anytorch.compile/ inductor node) compiles the model through Triton for another speedup.
⚠️ The risk. Both are version-locked to your exact torch + CUDA + python. A wrong wheel doesn't just fail to install — it can break the torch install (mismatched CUDA DLLs,
ImportError, or silent NaNs). And the failure mode of not having them is a hard crash before any sampling:ValueError: Can't import SageAttention: No module named 'sageattention', or compile errors /triton: unavailablein the startup log. This is exactly thevideo-extendTRAP 5.
✅ Therefore the default is: get a working render FIRST with the sdpa / no-compile fallback, then OFFER to install acceleration for speed. Never silently run a torch-breaking install to "fix" a workflow — fall back, render, then ask.
⚠️ Verification note (June 2026). Wheel sources, the triton↔torch table, and the live
attention_modeenum below were verified againstwoct0rdho/triton-windows,woct0rdho/SageAttentionreleases, and WanVideoWrapper's nodes (see Sources). Versions move fast — always re-read the live torch/CUDA/python first (commands below) and pick the wheel that matches; flag anything you can't confirm rather than guessing.
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.
- 10d ago First seen · 384 lines · 0 tokens per session scan C e199b6170b9a
triton-sageattention is a skill published in the GitHub repository sandyup/comfyui-mcp (1 stars, last pushed 2mo ago), licensed MIT. It adds 170 tokens to every session and 5,610 once invoked, about $0.0009 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). It is 86% identical to triton-sageattention, differing in 270 lines, and is treated as a copy.
Other skills, from other repositories
webgl-holographic-foil
A self-contained WebGL2 hero: thin-film interference over a crushed-foil surface whose palette shifts with the viewing angle; move the cursor to tilt the film.
general-video
Author or edit a custom HyperFrames composition when no specialized workflow fits, or when BRIEF.md sets flow: companion. Use for longer or multi-scene pieces, brand and sizzle reels, montages, static loops, static title cards, footage remixes, and freeform builds. Use motion-graphics instead for a short unnarrated…
html-ppt-hermes-cyber-terminal
OpenDesign + BYOK: choosing and wiring your own model, hands-on — cost, quality, and the routing decision. Built as a decision-grade AI literacy deck for engineers, IT, applied-AI teams.
html-ppt-taste-brutalist
16:9 HTML deck in tactical-telemetry / CRT-terminal taste. Deactivated-CRT charcoal slides, white-phosphor monospace, hazard-red accent, scanline overlay, ASCII syntax, density over decoration. Distilled from Leonxlnx/taste-skill brutalist-skill (Tactical Telemetry mode).
diagnostic-stem-delivery
Audio production with diagnostic analysis, timecode parsing from documents, and verified export workflow.
chengfeng-check-updates
An environment manager for a video-editing system. It checks whether its skills and runtime—the software needed to run them—are installed and compatible.