ワンクリックでManusで任意のスキルを実行

config-tuner

スター11

フォーク17

更新日2026年3月18日 22:16

Skill for helping Solutions Architects and Customer Engineers sweep configurations and find optimal end-to-end stack setups.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

llm-d

llm-d/llm-d-prism

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

Config Tuner / Solutions Architect Skill

User Profile Goal

The primary goal of this profile is to recommend optimal component-level and full end-to-end stack configurations (hardware infrastructure, model serving, orchestration) to customers based on rigorous testing.

Agent Responsibilities

When designing and developing features for this user profile, you must:

Orchestrate Complex Sweeps:
- Generate and run automated benchmarks across complex sweeps of configurations.
- Key metrics to sweep over: #P (Prompt size), #D (Decode size), Parallelism (TP, EP, DP), vLLM parameters, and generalized infrastructure configs (IGW, scheduler, KV cache).
Isolate Test Environments:
- Ensure each benchmark is run against a freshly deployed stack so that the configuration changes are completely isolated and not affected by previous state.
Visualize Relative Performance:
- Create tables, heatmaps, or charts that allow the user to immediately understand the relative performance impact of different configuration combinations.
Produce Reference Architectures:
- Synthesize the optimal benchmark results into clear, experimentally verifiable reference architectures that guarantee specific scalability Service Level Objectives (SLOs).

このリポジトリの他の Skills

同じリポジトリ

enforce-development-loop

llm-d/llm-d-prism

Enforce the Product -> UI/UX -> Engineering development loop for new features. Run when starting a feature, reviewing progress, or before merging to main.

2026-07-0911

benchmark-developer

llm-d/llm-d-prism

Skill for Analyst / PR roles to compare cross-stack performance and publish reproducible, community-facing benchmarks.

2026-03-1811

llm-d-feature-developer

llm-d/llm-d-prism

Skill for helping llm-d Feature Developers run, compare, and validate benchmarks for new optimization features.

2026-03-1811

llm-d-stack-operator

llm-d/llm-d-prism

Skill for helping Stack Operators monitor production stacks, perform regular stress tests, and detect regressions.

2026-03-1811

name	config-tuner
description	Skill for helping Solutions Architects and Customer Engineers sweep configurations and find optimal end-to-end stack setups.

Config Tuner / Solutions Architect Skill

User Profile Goal

Agent Responsibilities

When designing and developing features for this user profile, you must:

Orchestrate Complex Sweeps:
- Generate and run automated benchmarks across complex sweeps of configurations.
- Key metrics to sweep over: #P (Prompt size), #D (Decode size), Parallelism (TP, EP, DP), vLLM parameters, and generalized infrastructure configs (IGW, scheduler, KV cache).
Isolate Test Environments:
- Ensure each benchmark is run against a freshly deployed stack so that the configuration changes are completely isolated and not affected by previous state.
Visualize Relative Performance:
- Create tables, heatmaps, or charts that allow the user to immediately understand the relative performance impact of different configuration combinations.
Produce Reference Architectures:
- Synthesize the optimal benchmark results into clear, experimentally verifiable reference architectures that guarantee specific scalability Service Level Objectives (SLOs).