메뉴
Use when the user asks to install, launch, or first-run Bracket; or asks "how do I get started with Bracket". Walks through install.sh / install.ps1, launches the unified server on port 8000, and verifies the React UI loads.
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
Use when the user wants to add support for a new diffusion model or LLM training script in Bracket. Covers the Trainer protocol, search-space declaration, curated configs, presets, and tests. Lists every currently-supported preset so you know whether the model is already wired in.
Use when the user wants to create a dataset.toml for a Bracket / musubi-tuner / sd-scripts training session, or asks "how do I configure my dataset", "what does the toml file look like", "I have a folder of images, what's next". Walks through asking only the questions the toml format actually needs (paths, resolution, repeats), validates the result, and writes the file.
Use when the user wants to run a Bracket orchestration session — picking a trainer, dataset, budget, then kicking off the search. Covers both UI-driven and CLI-driven flows, including the interactive path where you collect missing inputs (paths, prompts) from the user one question at a time. Use when they say things like "run a Bracket sweep", "fine-tune Z-Image with Bracket", "search for the best LoRA config".
Use when a Bracket run is failing, hanging, producing no scores, only running for a handful of steps, or the dashboard is showing weird state. Walks through ledger inspection, log triage, the LM Studio VRAM-leak failure mode, and the most common other failure classes.
| name | bracket-quickstart |
| description | Use when the user asks to install, launch, or first-run Bracket; or asks "how do I get started with Bracket". Walks through install.sh / install.ps1, launches the unified server on port 8000, and verifies the React UI loads. |
When the user wants to install or launch Bracket, follow these steps in order.
install.sh and launch.shinstall.ps1 and launch.ps1 (or the .bat
wrappers install.bat / launch.bat)Run from the repo root:
./install.sh # Linux / macOS / WSL2
.\install.ps1 # Windows PowerShell
install.bat # Windows cmd.exe
The installer:
nvidia-smi and picks the matching PyTorch wheel
(cu124 / cu121 / cu118 / CPU)..venv/ and installs Bracket editable + dev deps.npm ci && npm run build to build the React
bundle into frontend/dist/ (warn-and-skip if Node missing).vendor/musubi-tuner and vendor/sd-scripts git submodules
pinned to specific commits in .gitmodules.vendor/venv/ with PyTorch +
trainer requirements + editable install of musubi-tuner.~/.lmstudio/bin/lms[.exe] and writes
BRACKET_LMS_BIN to .env (needed so the VLM judge can release
VRAM between runs — see bracket-debug-run)..env with sensible defaults.If install fails: most likely causes are git / Node / Python missing
from PATH, or the repo root not being writable. Re-running is safe —
every step is idempotent.
./launch.sh # Linux / macOS / WSL2
.\launch.ps1 # Windows PowerShell
launch.bat # Windows cmd.exe
This starts FastAPI (with the built React SPA mounted at /) on
http://127.0.0.1:8000. The browser doesn't auto-open — point at that
URL.
If the URL returns 404 instead of the UI: the React bundle isn't built.
Run cd frontend && npm run build then refresh, or re-run the
installer.
Once the UI is up:
LoRA or Full FT. Required fields appear (*
marked).vendor/ submodules — the user shouldn't have to type them.BRACKET_*_PATH env vars in .env.If the UI doesn't load: check the server logs (printed to stdout) for
import errors. Common: pandas or pydantic missing — pip install -e . from .venv usually fixes.
Bracket needs a base model checkpoint to fine-tune against. Each preset in the UI lists the env-var name in its help text. Common ones:
| Family | Env vars to populate |
|---|---|
| SDXL | BRACKET_SDXL_PRETRAINED |
| Z-Image | BRACKET_VAE_PATH, BRACKET_QWEN3_TE_PATH, plus DiT path per session |
| Flux-2-Klein | BRACKET_FLUX2_DIT_PATH, BRACKET_VAE_PATH, BRACKET_MISTRAL3_TE_PATH |
| Flux.1 / Flux.1-Kontext | BRACKET_FLUX1_DIT_PATH, BRACKET_FLUX1_AE_PATH, BRACKET_T5XXL_PATH, BRACKET_CLIP_L_PATH (+ BRACKET_FLUX1_KONTEXT_DIT_PATH for Kontext) |
| Qwen-Image / Qwen-Image-Edit | BRACKET_QWEN_IMAGE_DIT_PATH, BRACKET_QWEN_IMAGE_VAE_PATH, BRACKET_QWEN_IMAGE_TE_PATH (TE is Qwen2.5-VL-7B, NOT Qwen3) |
| SD3.5 | BRACKET_SD35_PRETRAINED (single file with MMDiT + T5 + CLIPs) |
| HunyuanVideo / FramePack | BRACKET_HUNYUAN_VIDEO_DIT_PATH, BRACKET_HUNYUAN_VIDEO_VAE_PATH, BRACKET_LLAMA3_PATH, BRACKET_CLIP_L_PATH |
| Wan 2.1 / 2.2 | BRACKET_WAN_DIT_PATH, BRACKET_WAN_VAE_PATH, BRACKET_UMT5_PATH |
| LTX-2 | BRACKET_LTX2_MODEL_PATH, BRACKET_LTX2_TEXT_ENCODER_PATH (native ltx-trainer) |
Don't bundle weights — they're 5-25 GB per family. Tell the user to
download the ones they need from Hugging Face / the model's release
page, then either set the env vars in .env or fill the path field in
the Setup tab once per session.
Hand off to the bracket-run-session skill. It walks through dataset
selection, judge configuration, budget, and stage-2 finals.
Hand off to the bracket-dataset-toml skill. It generates a valid
musubi/sd-scripts dataset config from a directory of images +
captions, asking the user only for the paths it can't infer.