Deploy_RoboorchardLab_Skills_to_Agent

[wuruize2002]

Deploy_RoboorchardLab_Skills.zip

[wuruize2002]

This issue targets users deploying RoboOrchardLab on Alibaba Cloud remote GPU servers (specifically bare-metal
H20/Hopper instances with no outbound internet). I encountered and documented 29 non-obvious deployment pitfalls,
organized into five categories as structured, machine-readable skill documents — each skill follows a Trigger → Root
Cause → Solution → Verification format, designed to be consumed directly by AI agents (e.g., Claude Code) so that an agent can autonomously diagnose and fix these issues during deployment. The repository also
includes a complete 12-phase deployment guide (Phase 0–11) as a standalone HTML page, which an agent can parse step by
step. The five categories are:

1. Environment Configuration — non-interactive shells skip .bashrc due to Ubuntu's PS1 guard; Alibaba Cloud VPC
requires an internal HTTP pip mirror; SCP through a bastion host needs the -O flag (OpenSSH 9.0+ defaults to SFTP);
pip --target mode has three fatal behaviors (ghost dist-info, no overwrite without --upgrade, and uncontrolled
dependency tree that can replace your torch); and nvcc --version (not nvidia-smi) must match torch.version.cuda.

2. PyTorch / Core / Lab Installation — offline wheel + SCP workflow; bare-metal LD_LIBRARY_PATH must cover 7
nvidia/triton subdirectories; make install (not pip install .) is required because setup.py reads a VERSION_POSTFIX
file generated by make version; robo_orchard_core must be installed at a specific git commit and requires a .git
directory; a missing robo_orchard_schemas stub breaks make install when ROBO_ORCHARD_USE_LOCAL=1; pip -c only accepts file paths, not inline strings; and --no-build-isolation is mandatory for 5 packages whose setup.py imports torch or
reads build-time generated files.

3. Version Conflicts — numpy<2 is a hard constraint but gets silently pulled to 2.x by torch, opencv, etc.;
opencv-python-headless>=4.12 declares numpy>=2, creating a direct conflict; pydantic must sit in the narrow window >=2.9.0, <=2.10.6; duckdb<=1.3.2; transformers is only compatible with torch 2.7.0 in the 4.49–4.51 range (5.x
references torch.nn.attention.flex_attention.AuxRequest which doesn't exist); huggingface-hub<1.0; and after every
extras install, you must pip uninstall opencv-python -y && pip install --force-reinstall
opencv-python-headless==4.11.0.86 --no-deps because nearly every extra re-introduces the GUI opencv.

4. Source Compilation — pytorch3d requires the full CUDA toolkit + gcc + ninja toolchain; holobrain_0 is missing from
the [all] extras list; robo_orchard_schemas needs a protobuf compile chain and its protobuf<6.0dev declaration is a
bug (generated code actually requires >=6.31.1); and foxglove-sdk has a different import name (foxglove) than its pip
package name.

5. Extras Implicit Dependencies — undeclared cross-extras import-time dependencies form a graph: bip3d ←
sem/holobrain, monodream ← progress_think, (monodream + progress_think + aux_think) ← mapdream; plus two external
implicit dependencies (llava ← aux_think, janus ← mapdream) that must be installed with --no-deps to avoid their
hard-pinned downstream version overrides.

The full repository includes all 29 skills and the 12-phase deployment guide, intended as a ready-to-use knowledge
base for both human operators and AI coding agents.

[AlanHuangT]

如果是内部开发人员， 走内部的merge request 。 如果是外部开发人员，暂时还不支持外部合入。