multi-gpu-dispatch
/multi-gpu-dispatch
Rozkłada produkcję na kilka kart GPU lub serwerów ComfyUI: rozdziela ujęcia po serwerach albo dzieli duży model między karty.
- Autor
- mojomast ↗
- Repozytorium
- open-higgsfield-skills ↗
- Licencja
- MIT
- Ostatnia zmiana repozytorium
- 30.10.2026
Ocena i bezpieczeństwo
- Pliki w paczce
- 15
- Skrypty
- 3
- Komentarze HTML
- 0
- Adresy URL w SKILL.md
- 2
Brak sygnałów. Repozytorium zmienione 30.10.2026, bez odwołań do wycofanych modeli i API.
MIT, otwarta: można używać i zmieniać, z zachowaniem informacji o autorze.
- 3 skrypty w paczce. Przejrzyj je przed instalacją, bo Claude może je uruchomić.
Skrypty: scripts/comfy_client.py, scripts/dispatch.py, scripts/preflight.py
Co robi i kiedy
Rozkłada produkcję na kilka kart GPU lub serwerów ComfyUI: rozdziela ujęcia po serwerach albo dzieli duży model między karty.
Oryginalny opis (EN)
Spread a production across several GPUs or ComfyUI servers: round-robin shots across server URLs with a stdlib dispatcher, or install ComfyUI-Distributed (parallel workers and tiles) or raylight (split one large model across GPUs with FSDP/xDiT). Use when one GPU is too slow or too small for the plan.
Jak działa
- dispatch.py
- ComfyUI-Distributed and raylight
- Rules
Podgląd SKILL.md
--- name: multi-gpu-dispatch description: Spread a production across several GPUs or ComfyUI servers: round-robin shots across server URLs with a stdlib dispatcher, or install ComfyUI-Distributed (parallel workers and tiles) or raylight (split one large model across GPUs with FSDP/xDiT). Use when one GPU is too slow or too small for the plan. --- # multi-gpu-dispatch There are three levels. Pick the lowest one that solves the problem. | need | use | license | |---|---|---| | many independent shots, several servers (local, LAN, Google Colab tunnels, cloud) | `scripts/dispatch.py` (original, stdlib) | this collection | | one graph, run in parallel on several GPUs (seeds and batches, tiled upscale, Wan t2v) | **robertvoy/ComfyUI-Distributed** node pack | Apache-2.0 | | one model too big for one GPU (Wan 2.2 14B, LTX-2.3 22B, Flux) | **komikndr/raylight** node pack (FSDP / xDiT USP) | Apache-2.0 | ## dispatch.py `python scripts/dispatch.py plan manifest.json --url http://gpu0:8188 --url http://gpu1:8188 [--weights 2,1]` - It assigns each pending shot to a server. Weighted round-robin is used, chained shots (`chain_from`) stay on the server of the shot they chain from, and the assignment is written to `<manifest>.dispatch.json`. - `run --dry-run` prints the per-server queue without network access. - `run` writes one manifest copy per server and starts `video-director/scripts/direct.py run <copy> --only s01,s03…` for each one in parallel. Each worker does stills and video for its own shots only, and each server has to pass its own preflight. Approve the `bible`, `plan` and `spend` gates on the main manifest **before** you dispatch, because the copies inherit them. - `merge` copies the worker shots back. Then run `direct.py run manifest.json` once for continuity, audio, post and assembly. - Budget: the GPU-minute caps are per manifest, summed across servers. ## ComfyUI-Distributed and raylight Their READMEs are copied in `references/` (Apache-2.0) and cover install steps, worker setup and limits. Example graphs, all UI format: - `ui/distributed_wan22_14b_t2v.ui.json` - `ui/distributed_upscale_video.ui.json` - `ui/raylight_ltx23_i2v.ui.json` - `ui/raylight_wan22.ui.json` `workflows/ui/*.ui.json` are official Comfy-Org templates (MIT) in **UI format**, and most contain subgraphs. `comfy_client.py` refuses UI format. To use one: open it in ComfyUI, install the missing nodes and models that ComfyUI lists (they are also in `requirements.md`), then use **Workflow → Export (API)** to save it as `workflows/<name>.api.json`. Run `preflight.py` on the export and read off the node ids for `--set`. Node ids change when you export, so do not reuse ids from another graph. ## Rules - The user installs node packs on their own servers. Never install packs or start workers from this skill. - Every remote worker on Google Colab or Comfy Cloud spends credit. Show the per-worker estimate, then pass the video-director spend gate. - raylight needs matched GPUs and NCCL. Distributed needs the same models on every worker. Check with `preflight.py` against **each** URL.
Pliki w paczce (15)
- SKILL.md 3 KB
- references/ComfyUI-Distributed-README.md 13 KB
- references/raylight-README.md 19 KB
- references/requirements.md 9 KB
- scripts/comfy_client.py 8 KB
- scripts/dispatch.py 4 KB
- scripts/preflight.py 3 KB
- workflows/api/distributed_upscale_video.api.json 4 KB
- workflows/api/distributed_wan22_14b_t2v.api.json 6 KB
- workflows/api/raylight_ltx23_i2v.api.json 12 KB
- workflows/api/raylight_wan22.api.json 6 KB
- workflows/ui/distributed_upscale_video.ui.json 17 KB
- workflows/ui/distributed_wan22_14b_t2v.ui.json 19 KB
- workflows/ui/raylight_ltx23_i2v.ui.json 59 KB
- workflows/ui/raylight_wan22.ui.json 20 KB
Wersje
W bazie nie ma innych skilli o nazwie multi-gpu-dispatch.
Migawka bazy z 13.11.2026 · wersja paczki 65a3f9eb96be (skrót SHA-256 całej paczki)
Pobrania w ostatnich 30 dniach: 0 · ranking
Zgłoś problem z tym skillem: szkodliwa treść, błędny opis albo prośba autora o usunięcie.
Podobne skille
- 100comfyui
Generuje obrazy i wideo AI przez workflow ComfyUI (FLUX, Stable Diffusion, modele wideo) na lokalnym GPU lub rozproszonej mocy Salad Cloud. - 93vram-radar
Pokazuje, które procesy zajmują kartę graficzną i VRAM, czy inny model się zmieści i co zwolnić, także dla ComfyUI i Ollama. - 90gpu-cluster-scheduling
Szereguje zadania GPU na współdzielonym klastrze tak, by zadania rozproszone dostawały wszystkie GPU naraz, fragmentacja była niska, a wywłaszczanie przewidywalne. - 100framefields-cloud
Renderuje projekty framefields na chmurowych GPU lub własnym workerze, planuje koszt i dzieli długi film między kilka GPU. - 96gpu-use
Sprawdza na zdalnym serwerze zużycie GPU: pamięć każdej karty, działające procesy i należące do nich kontenery. - 100cloudgpu-comfyui-images
Wynajmuje GPU do generowania obrazów w ComfyUI modelem FLUX, gdy lokalna karta nie wystarcza.