界面已本地化;技能介绍与说明文档按所选语言由 AI 自动翻译,代码、命令、路径和专业术语保留原文,以原作者文档为准。
探索技能
效率与协作 · SKILL.md

environment-and-model-setup

environment-and-model-setup

【效率与协作】Set up and verify Hunyuan3D-2 Python, CUDA, extension, and model-cache environments.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: environment-and-model-setup description: "Set up and verify Hunyuan3D-2 Python, CUDA, extension, and model-cache environments." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Hunyuan3D-2 Environment and Model Setup

Use this sub-skill when the task is about installing hy3dgen, choosing PyTorch/CUDA variants, building texture extensions, validating GPU/backend readiness, selecting model subfolders, or planning local/offline model caches.

Route here for

  • Installing package requirements and hy3dgen from a package source or a repository tree the user is intentionally working in.
  • Building custom_rasterizer and mesh_processor for texture generation.
  • Diagnosing CUDA toolkit/header, PyTorch CUDA, pymeshlab/OpenGL, and extension import failures.
  • Understanding model zoo names, subfolders, VRAM needs, and HY3DGEN_MODELS cache behavior.
  • Running non-downloading environment checks before shape/texture/service workflows.

Do not route here for

  • Shape generation code parameters after the environment is ready: use ../shape-generation/.
  • Texture/mesh workflow details after extensions are ready: use ../texture-and-mesh/.
  • API server/Gradio/Blender client behavior: use ../services-and-integrations/.

Essential references

  • Installation and backends for installation order, CUDA extension build guidance, and backend readiness checks.
  • Model overview for model repo/subfolder selection, VRAM guidance, and cache behavior.
  • Troubleshooting for common install/backend failures and stale documentation warnings.

Bundled helper

  • scripts/check_install.py checks package metadata/imports and optional CUDA/extension readiness without downloading model weights.

Examples:

python scripts/check_install.py --json
python scripts/check_install.py --check-cuda --json
python scripts/check_install.py --check-cuda --check-extensions --json

A passing helper check proves installation/backend readiness for the selected probes. It does not prove that full model inference ran or that model checkpoints are cached.

我的收藏