reinvent4（reinvent4 效率）

REINVENT4

Prefer a local REINVENT4 checkout over web summaries. Treat a directory as the repo root when it contains install.py, pyproject.toml, reinvent/, and configs/.

Workflow

1. Classify the request: install, run configuration, data preprocessing,

notebook conversion, scoring plugin, or test/troubleshooting.

1. Read only the relevant reference file:

- installation or CLI usage: references/install-and-run.md - TOML mode selection or parameter mapping: references/config-modes.md - plugins, notebooks, or tests: references/plugins-and-tests.md

1. Verify commands against the local checkout before proposing them. Prefer

python install.py --help, python install.py <backend> --dry-run, reinvent --help, direct file inspection, and existing files under configs/.

1. Reuse the example configs in configs/ instead of inventing schemas from

scratch.

1. Keep file paths explicit. Upstream example configs are templates and must be

adjusted to local model, SMILES, output, and log paths before execution.

Installation Rules

• Use an isolated Python environment with Python 3.10 or newer.
• Map the processor/backend carefully:

- macOS CPU: mac - Linux CPU: cpu - Linux NVIDIA CUDA: upstream examples use values like cu126 - Linux AMD ROCm: upstream examples use values like rocm6.4 - Intel XPU: xpu - Windows: CPU, CUDA, and XPU are supported, but upstream says Windows is only partially tested

• Remember that install.py defaults to optional dependency set all, which

includes extra packages such as openeye and isim.

• Prefer -d none for minimal or smoke-test installs unless the user

explicitly needs OpenEye ROCS or iSIM-related functionality.

• Use --dry-run before a real install whenever backend choice or dependency

resolution is uncertain.

• Verify a finished install with reinvent --help.

Running REINVENT

• Main CLI entry point: reinvent [-l logfile] <config.toml>.
• Data pipeline entry point from pyproject.toml: reinvent_datapre.
• Prefer TOML because upstream ships maintained examples in configs/.
• When editing configs, update at least device selection, model/prior paths,

SMILES inputs, output files, and TensorBoard/log paths.

Troubleshooting Rules

• On macOS, remind the user that upstream documents CPU-only support and says

macOS is only partially tested.

• If a macOS clone reports path collisions under

contrib/tutorials/maize/adgpu_prepare, avoid relying on those collided files unless the user specifically needs that tutorial.

• For tests, warn that they require a JSON config with a non-existent

MAIN_TEST_PATH; some tests also require OE_LICENSE.

• Do not promise a full RL/TL run unless models, datasets, and optional

licensed tools are present locally.

Source Files

Use these upstream files as the primary source of truth when they are present in the local checkout:

• README.md
• install.py
• pyproject.toml
• configs/README.md
• configs/PARAMS.md
• configs/SCORING.md
• notebooks/README.md
• tests/example_config.json