메뉴
AI-Toolkit backend and image workflow for Kura. Use when changing AI-Toolkit compile output, Docker image build/publish, Hugging Face cache behavior, or local/RunPod image contracts.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
ComfyUI render workflow operations for Kura. Use when editing workflows/*.json, render run.yaml files, workflow_patches, promptsets, image comparison generation, ComfyUI endpoint behavior, or render result collection.
RunPod remote training lifecycle and billing safety for Kura. Use when working on kura run remote, runpod staging/upload/download/pull/stop/reconcile, Pod cleanup, max lease, notifications, GPU selection, Network Volumes, or RunPod docs/README updates.
Dataset preparation and validation for Kura training runs. Use when creating or editing datasets/, dataset.yaml, items.jsonl, captions, trigger words, paired/control datasets, dataset roles, or dataset validation behavior.
Local disk, WSL2, Docker, cache, cleanup, permission, and checkpoint safety workflow for Kura. Use before local Docker training, real model-download smoke tests, Docker image builds, RunPod downloads/pulls, cleanup, permission repair, or when the user mentions disk space, WSL, Docker storage, cache, checkpoints, root-owned files, cleanup, or a crash/reconnect incident.
Musubi Tuner backend guidance for Kura. Use when compiling or changing Musubi compile output, model bundle resolution, safetensors validation, LoRA output compatibility, dataset TOML generation, low-VRAM flags, batch/gradient accumulation, or Musubi Docker images.
Choosing training parameters and resource trade-offs for Kura LoRA runs. Use when proposing or reviewing run.yaml training parameters, judging VRAM fit from `kura run plan` Resources, reacting to CUDA OOM, choosing fp8/quantized artifacts, or deciding which memory-saving or quality trade-off to apply and what needs user approval.
| name | ai-toolkit-backend |
| description | AI-Toolkit backend and image workflow for Kura. Use when changing AI-Toolkit compile output, Docker image build/publish, Hugging Face cache behavior, or local/RunPod image contracts. |
Use this skill for AI-Toolkit-specific backend and image work.
Use this ladder when logs or doctor output show the run does not fit available
VRAM. The agent may choose recipe-preserving execution accommodations while
drafting, but must expose them in the plan and record them in run.yaml before
recompiling. Ask separately when the change affects the training recipe, GPU
cost, or is expected to increase elapsed time beyond roughly 2x.
quantize,
quantize_te, and backend-supported low_vram options.run.yaml.uv run kura run plan <run-id> and get
explicit approval.uv run kura image build ai-toolkit --ref <branch-or-commit>
uv run kura image inspect ai-toolkit
uv run kura image publish ai-toolkit --dry-run
uv run kura doctor docker
uv run kura run compile <run-id>
uv run kura run launch <run-id> --executor docker --dry-run
README may lag implementation. Verify with CLI help before documenting flags.