快速模型执行控制器
AI Skill Hub 推荐使用：快速模型执行控制器 是一款优质的AI工具。AI 综合评分 7.5 分，在同类工具中表现稳健。如果你正在寻找可靠的AI工具解决方案，这是一个值得深入了解的选择。
📚 深度解析
**为什么要使用开源工具而非商业 SaaS？**
对于个人开发者和有隐私需求的用户，本地部署的开源工具意味着数据不离本机，不受第三方服务商的数据政策约束。同时，开源工具通常没有使用次数限制和月度费用，一次安装即可长期使用，对于高频使用场景的总拥有成本（TCO）远低于订阅制商业工具。
**安装与环境准备**
快速模型执行控制器 依赖 Go 运行环境。建议通过 pyenv（Python）或 nvm（Node.js）管理 Go 版本，避免全局环境污染。对于新手用户，推荐先创建虚拟环境（python -m venv venv && source venv/bin/activate），再安装依赖，这样即使出现问题也可以随时删除虚拟环境重新开始，不影响系统稳定性。
**社区与维护**
GitHub Issue 和 Discussion 是获取帮助的最快渠道。在提问前建议先检查 Closed Issues（已关闭的问题），大多数常见问题都已有解答。遇到 Bug 时，提供 pip list 的输出、完整错误堆栈和最小可复现示例，能显著提高开发者响应速度。AI Skill Hub 将持续追踪 快速模型执行控制器 的版本更新，及时通知重要功能变化。
📋 工具概览
使用vLLM睡眠/唤醒和启动实现快速模型执行
快速模型执行控制器 是一款基于 Go 开发的开源工具，专注于 llm-d、go、kubernetes 等核心功能。作为 GitHub 开源项目，它拥有活跃的社区支持和持续的版本迭代，代码完全透明可审计，支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流，都能提供稳定可靠的解决方案。
📖 中文文档
使用vLLM睡眠/唤醒和启动实现快速模型执行
快速模型执行控制器 是一款基于 Go 开发的开源工具，专注于 llm-d、go、kubernetes 等核心功能。作为 GitHub 开源项目，它拥有活跃的社区支持和持续的版本迭代，代码完全透明可审计，支持本地部署以保护数据隐私。无论是个人使用还是集成到企业工作流，都能提供稳定可靠的解决方案。
- 开源免费，支持本地部署，数据完全自主可控
- 活跃的 GitHub 开源社区，持续迭代更新
- 提供详细文档和使用示例，新手友好
- 支持自定义配置，灵活适配不同使用环境
- 可作为基础组件集成进现有技术栈或进行二次开发
- 本地部署运行，保护数据隐私，满足合规要求
- 自定义集成到现有系统，扩展技术栈能力
- 作为开源基础组件进行商业化二次开发
# 方式一：go install（推荐） go install github.com/llm-d-incubation/llm-d-fast-model-actuation@latest # 方式二：从源码编译 git clone https://github.com/llm-d-incubation/llm-d-fast-model-actuation cd llm-d-fast-model-actuation go build -o llm-d-fast-model-actuation . # 方式三：下载预编译二进制 # 访问 Releases 页面下载对应平台二进制文件 # https://github.com/llm-d-incubation/llm-d-fast-model-actuation/releases
- 访问 GitHub 仓库页面
- 按照 README 文档完成依赖安装
- 根据系统环境完成初始化配置
- 参考官方示例或文档开始使用
- 遇到问题可在 GitHub Issues 中查找解答
# 查看帮助 llm-d-fast-model-actuation --help # 基本运行 llm-d-fast-model-actuation [options] <input> # 详细使用说明请查阅文档 # https://github.com/llm-d-incubation/llm-d-fast-model-actuation
# llm-d-fast-model-actuation 配置说明 # 查看配置选项 llm-d-fast-model-actuation --config-example > config.yml # 常见配置项 # output_dir: ./output # log_level: info # workers: 4 # 环境变量（覆盖配置文件） export LLM_D_FAST_MODEL_ACTUATION_CONFIG="/path/to/config.yml"
简介
The llm-d-fast-model-actuation repository is part of the llm-d ecosystem for serving large language models on Kubernetes. FMA lives in the llm-d-incubation organization, where new llm-d components are developed before graduation.
This repository contains work on one of the many areas of work that contribute to fast model actuation. This area concerns exploiting techniques in which an inference server process dramatically changes its properties and behavior over time.
There are two sorts of changes contemplated here. Both are currently realized only for vLLM and nvidia's GPU operator, but we hope that these ideas can generalize.
1. vLLM level 1 sleep and wake_up. A vLLM instance in level 1 sleep has its model tensors in main (CPU) memory rather than accelerator (GPU) memory. While in this state, this instance can not serve inference requests --- but has freed up accelerator resources for use by a different instance. But the sleeping instance is still a running process (e.g., it can still serve administrative requests) as far as the OS is concerned. The process of waking up the sleeping instance is very fast; for example, taking about 3 seconds for a model with 64 GiB of tensor data. This behavior is available in vLLM today.
2. Model swapping. In model swapping techniques, there is a persistent management process that can run various subsidiary inference server processes over time. The management process does basic code loading and initialization work of the inference server so that this work does not have to be done at the startup of the inference server process, reducing that startup latency. The inference servers may be able to sleep and wake up.
A process with such flexibility does not easily fit into the Kubernetes milieu. The most obvious and natural way in Kubernetes to define a desired inference server is to create a Pod object. However, a Pod has a static allocation of accelerator resources and a static command line. That is, the obvious way to define a Pod is such that it serves one fixed model and server options, with no resource-freeing hiatus. This repository contains a way of fitting the process flexibility into the Kubernetes milieu. We call this technique "dual pods". It makes a distinction between (a) a server-requesting Pod, which describes a desired inference server but does not actually run it, and (b) a server-providing Pod, which actually runs the inference server(s).
The topics above are realized by the following software components.
- A dual-pods controller, which manages the server-providing Pods in reaction to the server-requesting Pods that other manager(s) create and delete. This controller is written in the Go programming language and this repository's contents follow the usual conventions for one containing Go code.
- A vLLM instance launcher, the persistent management process mentioned above. This is written in Python and the source code is in the inference_server/launcher directory.
- A launcher-populator controller, which watches LauncherConfig and LauncherPopulationPolicy custom resources and ensures that the right number of launcher pods exist on each node. This controller is also written in Go.
These controllers are deployed together via a unified Helm chart at charts/fma-controllers. The chart also installs the shared RBAC resources and optional ValidatingAdmissionPolicies.
The repository defines three Custom Resource Definitions (CRDs):
- InferenceServerConfig — declares the properties of an inference server (image, command, resources) that server-providing Pods use. - LauncherConfig — declares the configuration for a launcher process (image, resources, ports) that manages vLLM instances. - LauncherPopulationPolicy — declares the desired population of launcher pods per node.
These CRD definitions live in config/crd and the Go types are in pkg/api.
The development roadmap has three milestones. Milestone 2, which introduced vLLM sleep/wake without the launcher, is finished. Milestone 3, which adds launcher-based model swapping where a persistent launcher process manages vLLM instances on each node, is under implementation.
For further design documentation, see the docs directory.
高质量的Kubernetes控制器，快速模型执行
- 需要 llm-d-fast-model-actuation 解决具体问题的开发者与运营人员
- 生产部署优先使用 Docker Compose 隔离依赖，并挂载 volume 持久化数据
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- 容器内无法访问宿主机 localhost — 使用 host.docker.internal
- Docker：llm-d-fast-model-actuation 提供官方镜像，docker compose up 一键启动
- CLI：直接 npm install -g / pip install，命令行调用
- 云端托管：可放在 Vercel / Railway / Fly.io 等 PaaS 平台
⚡ 核心功能
- 开源免费，支持本地部署，数据完全自主可控
- 活跃的 GitHub 开源社区，持续迭代更新
- 提供详细文档和使用示例，新手友好
- 支持自定义配置，灵活适配不同使用环境
- 可作为基础组件集成进现有技术栈或进行二次开发
- 需要 llm-d-fast-model-actuation 解决具体问题的开发者与运营人员
- 生产部署优先使用 Docker Compose 隔离依赖，并挂载 volume 持久化数据
- API key 直接提交到 git 仓库（请用 .env 并加入 .gitignore）
- 容器内无法访问宿主机 localhost — 使用 host.docker.internal
👥 适合人群
🎯 使用场景
- 本地部署运行，保护数据隐私，满足合规要求
- 自定义集成到现有系统，扩展技术栈能力
- 作为开源基础组件进行商业化二次开发
⚖️ 优点与不足
- +Apache-2.0 协议，可免费商用
- +完全开源免费，无授权费用
- +本地部署，数据完全自主可控
- +开发者社区支持，遇问题可查可问
- −安装和初始配置可能需要一定技术基础
- −功能完整性通常不如成熟商业产品
- −技术支持主要依赖开源社区，响应速度不稳定
AI Skill Hub 为第三方内容聚合平台，本页面信息基于公开数据整理，不对工具功能和质量作任何法律背书。
建议在沙箱或测试环境中充分验证后，再部署至生产环境，并做好必要的安全评估。
✅ Apache 2.0 — 宽松开源协议，可商用，需保留版权声明和 NOTICE 文件，含专利授权条款。
🔗 相关工具推荐
❓ 常见问题 FAQ
总体来看，快速模型执行控制器 是一款质量良好的AI工具，在同类工具中具备一定竞争力。AI Skill Hub 将持续追踪其更新动态，建议收藏备用，结合自身场景选择合适时机引入使用。
| 原始名称 | llm-d-fast-model-actuation |
| 原始描述 | 开源AI工具：Kubernetes controllers for fast model actuation using vLLM sleep/wake and launch。⭐16 · Go |
| Topics | llm-dgokubernetes |
| GitHub | https://github.com/llm-d-incubation/llm-d-fast-model-actuation |
| License | Apache-2.0 |
| 语言 | Go |
收录时间：2026-06-02 · 更新时间：2026-06-02 · License：Apache-2.0 · AI Skill Hub 不对第三方内容的准确性作法律背书。