DGX Spark One-Click Deploy for MiniMax H3 Video Generation
Open-source deployment package for NVIDIA DGX Spark (GB10) — run MiniMax H3 video generation with one command. Integrates keys-heretic Sol-Attention acceleration, 12 pre-tuned workflows.
Repo: alexlu0912/dgxspark_comfyui_minimax_h3
Why this project?
MiniMax H3 is one of the strongest open-source video models. Deploying on DGX Spark has challenges: model weights scattered across HuggingFace & ModelScope (~165 GB), Sol-Attention needs Blackwell (sm_121) adaptation, GB10 unified memory requires careful tuning. This project bundles all learnings.
Key Features
- Two deployment methods: From scratch (install wizard with progress bars) or clone from existing (~5 min)
- 12 pre-tuned workflows: Text-to-video, image-to-video, multi-shot, keyframes, full pipeline
- Sol-Attention acceleration: Pre-patched Blackwell Sol-Attn + Triton + batch VAE, ~6 min for 5s 720p video
- Full docs: README, WORKFLOWS.md, install_wizard.py with resume support
Quick Start
git clone https://gitee.com/alexlu0912_admin/dgxspark_comfyui_minimax_h3.git
cd dgxspark_comfyui_minimax_h3
python install_wizard.py
# Open browser → http://IP:8188 → drag workflow JSON
Built and tested on Ubuntu 24.04 aarch64, CUDA 13.0+, GB10 (sm_121). Disk >= 200 GB. PRs and issues welcome!