mxl-core

星标9

分支0

更新时间2026年2月14日 20:41

This skill should be used when users need the canonical MaxwellLink workflow, including socket vs embedded coupling, units/time handling, and MPI/HPC invariants.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

TaoELi

TaoELi/FermiLink

打开 GitHub 仓库查看创作者相关仓库

下载

MaxwellLink core workflow

Enforce invariants

Create new runnable inputs only under projects/YYYY-MM-DD-NAME/.
Keep runs reproducible by storing parameters in a config.json (or similar) and avoiding hard-coded absolute paths.
Never pass both hub=... and driver=... to maxwelllink.Molecule; provide exactly one.

Choose coupling mode

Use embedded mode when the driver can run in the same Python process:
- Construct molecules with Molecule(driver="<id>", driver_kwargs={...}).
Use socket mode when drivers must run as separate processes or nodes:
- Construct a SocketHub(...), pass hub=hub into each Molecule(...), and launch one mxl_driver --model <id> ... per molecule.
If local socket runs fail with MPI/network permission errors, prefer embedded mode (Molecule(driver="...", ...)) on local machines and reserve socket mode for SLURM/HPC.
For local socket mode, you can disable MPI auto-init with MPI4PY_RC_INITIALIZE=0 and ensure loopback NIC usage (use 127.0.0.1; set site-specific NIC selection env vars if MPI/OFI tries a non-loopback interface).

Handle units/time

For Meep: set time_units_fs on MeepSimulation(...) and keep it consistent when interpreting driver times.
For SingleMode/LaserDriven: work directly in atomic units and set dt_au explicitly.

Handle MPI and multi-process execution

Launch socket drivers only once (rank 0) under MPI; rely on MaxwellLink broadcasts for other ranks.
Set generous timeout for HPC and tune latency to match driver cost.
If attaching many molecules, consider store_additional_data=False for most molecules to limit memory usage; keep it True only for the molecules you intend to export/plot.

Validate before running

Use skills/mxl-project-scaffold/scripts/mxl_validate_project.py to sanity-check config.json and required files.
When debugging behavior, compare to patterns in tests/ and docs/source/usage.rst.

References

Solver choice: skills/mxl-core/references/solver_quickref.md
Driver choice: skills/mxl-core/references/driver_quickref.md

同仓库更多 Skills

同仓库

scientific-simulation-guide

TaoELi/FermiLink

Guide users through setting up and running scientific simulations with FermiLink after `pip install fermilink` and `fermilink init`. Covers package installation, provider setup, execution modes, HPC configuration, and workflow selection.

2026-05-179

optimize-report

TaoELi/FermiLink

Generate a Sphinx-ready report bundle (plots, RST, contract, per-accepted-commit pages) from a completed `fermilink optimize` workspace. Use when the user asks to visualize, export, or publish optimization results from a `.fermilink-optimize` directory, or wants to prepare material for the Project Optimization website.

2026-04-229

optimize-goal-authoring

TaoELi/FermiLink

Write, review, or strengthen `goal.md` files for `fermilink optimize` across Python, C/C++, and Fortran packages. Use when defining target scope, scientific correctness constraints, representative workloads, build commands, or staged workload-input files for goal mode.

2026-04-119

mxl-project-scaffold

TaoELi/FermiLink

This skill should be used when users ask to prepare MaxwellLink input files; it scaffolds a new `projects/YYYY-MM-DD-NAME/` folder from copy-ready templates and provides basic validation.

2026-02-149

sci-skills-generator

TaoELi/FermiLink

This skill should be used when users ask to bootstrap or regenerate a `skills/` folder for a scientific software package (for example LAMMPS, PSI4, GROMACS, Q-Chem docs-only, MEEP, or similar) from docs/source/tests/tutorial layouts, with documentation-first routing and an abstract skill count capped around 30.

2026-02-149

mxl-hpc-slurm

TaoELi/FermiLink

This skill should be used when users need to run MaxwellLink in SLURM/HPC environments, including dependent job patterns and robust socket host/port handoff.

2026-02-139

MaxwellLink core workflow

Enforce invariants

Create new runnable inputs only under projects/YYYY-MM-DD-NAME/.

Keep runs reproducible by storing parameters in a config.json (or similar) and avoiding hard-coded absolute paths.

Never pass both hub=... and driver=... to maxwelllink.Molecule; provide exactly one.

Choose coupling mode

Use embedded mode when the driver can run in the same Python process:

Construct molecules with Molecule(driver="<id>", driver_kwargs={...}).

Use socket mode when drivers must run as separate processes or nodes:

Construct a SocketHub(...), pass hub=hub into each Molecule(...), and launch one mxl_driver --model <id> ... per molecule.

If local socket runs fail with MPI/network permission errors, prefer embedded mode (Molecule(driver="...", ...)) on local machines and reserve socket mode for SLURM/HPC.

For local socket mode, you can disable MPI auto-init with MPI4PY_RC_INITIALIZE=0 and ensure loopback NIC usage (use 127.0.0.1; set site-specific NIC selection env vars if MPI/OFI tries a non-loopback interface).

Handle units/time

For Meep: set time_units_fs on MeepSimulation(...) and keep it consistent when interpreting driver times.

For SingleMode/LaserDriven: work directly in atomic units and set dt_au explicitly.

Handle MPI and multi-process execution

Launch socket drivers only once (rank 0) under MPI; rely on MaxwellLink broadcasts for other ranks.

Set generous timeout for HPC and tune latency to match driver cost.

If attaching many molecules, consider store_additional_data=False for most molecules to limit memory usage; keep it True only for the molecules you intend to export/plot.

Validate before running

Use skills/mxl-project-scaffold/scripts/mxl_validate_project.py to sanity-check config.json and required files.

When debugging behavior, compare to patterns in tests/ and docs/source/usage.rst.

References

Solver choice: skills/mxl-core/references/solver_quickref.md

Driver choice: skills/mxl-core/references/driver_quickref.md

name	mxl-core
description	This skill should be used when users need the canonical MaxwellLink workflow, including socket vs embedded coupling, units/time handling, and MPI/HPC invariants.