인터페이스는 현지화되어 있으며, Skill 소개와 설명 문서는 선택한 언어로 AI가 자동 번역합니다. 코드·명령어·경로·전문 용어는 원문을 유지하며 원작자 문서가 우선합니다.
Skills 둘러보기
생산성 및 협업 · SKILL.md

extension-and-evolution

extension-and-evolution

Extend mergekit with custom merge methods and task graphs, or plan optional evolutionary merge searches with explicit dependency and evaluation gates.

Independent GitHub snapshot · Automated classification · Original documentation governs

Skill 문서

GitHub

name: extension-and-evolution description: "Extend mergekit with custom merge methods and task graphs, or plan optional evolutionary merge searches with explicit dependency and evaluation gates." disable-model-invocation: true metadata: disco-role: operating license: LGPL 3.0

Extension and evolution route

Use this route for a custom merge method, a reusable task-graph component, or mergekit-evolve. It is an extension/developer route, not the normal YAML method-selection route.

Route first

  • Use merge-configs for ordinary merge YAML, registered method selection, parameters, tokenizer, and chat templates.
  • Use model-io-and-architecture for model references, architecture conversion, tensor IO, devices, and memory.
  • Use specialized-workflows for raw, MoE, multi-stage, LoRA, tokensurgeon, or layer-shuffle commands.

Custom method procedure

  1. Define the mathematical contract and required parameters before choosing the decorator/easy-define or class-based implementation.
  2. Implement a tensor-local method that respects mergekit's task and device contract; register it under the exact YAML method name.
  3. Check registration in a clean import, exercise a tiny tensor fixture, and add focused tests for defaults, parameter validation, dtype, and device.
  4. Run the graph tests and a focused package test before using the method in a real model merge. Keep the source method self-contained and documented.

Read custom-methods.md for the API patterns and troubleshooting.md for signature, registry, and task-graph failures.

Evolution procedure

Treat mergekit-evolve as optional and expensive. Preflight the evolve extra, CMA-ES, evaluation backend, model/task config, storage, GPUs, and credentials before constructing a run. The inspected environment intentionally omits ray, cma, lm_eval, wandb, and vllm; mergekit-evolve --help therefore fails at import with a missing cma module. This is an explicit optional gate, not a core mergekit import failure.

Read evolution.md only after the preflight. Never start a networked evaluation, W&B run, Ray cluster, vLLM server, or long CMA-ES search as a smoke test. Record seeds, max evaluations, output paths, evaluator, and stop conditions.

Safe helper

Run python scripts/check_extension_registration.py --help or the default probe to inspect core registration without invoking optional evolution.

저장한 Skills