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

runtime-ops

runtime-ops

【效率与协作】Route Det3D installation, import, CUDA extension, PyTorch ABI, optional dependency, distributed-runtime, and checkpoint filesystem troubleshooting.

独立 GitHub 快照 · 自动分类 · 以原作者文档为准

技能说明

GitHub

name: runtime-ops description: "Route Det3D installation, import, CUDA extension, PyTorch ABI, optional dependency, distributed-runtime, and checkpoint filesystem troubleshooting." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Runtime Operations

Use this route whenever Det3D fails during installation, package import, CUDA operator loading/building, distributed initialization, or filesystem setup. Run scripts/check_runtime.py first; it is read-only and reports JSON.

Workflow

  1. Read installation.md for historical version constraints and installation order.
  2. Run the runtime diagnostic and separate framework CUDA availability from nvcc, compiler, extension, and spconv availability.
  3. Read custom-ops.md before rebuilding any extension; match Python, torch, CUDA toolkit, compiler, and GPU ABI.
  4. Check runtime-diagnostics.md for import classification and distributed environment checks.
  5. Apply troubleshooting.md and preserve the first actionable error.

Do not call a CPU import a successful GPU capability check. Do not repair a user-owned environment by silently changing torch, setuptools, CUDA, or compiled packages. Training/evaluation and data conversion remain separate side-effectful workflows.

我的收藏