Skip to main content

sageLLM Control Plane - Intelligent request routing, scheduling, and engine lifecycle management

Project description

sageLLM Control Plane

Protocol Compliance (Mandatory)

  • MUST follow Protocol v0.1: Protocol v0.1
  • Any globally shared definitions (fields, error codes, metrics, IDs, schemas) MUST be added to Protocol first.

CI Status PyPI version Python Versions License Code style: ruff

Intelligent request routing, scheduling, and engine lifecycle management for sageLLM.

职责定位

Control Plane 位于用户/Gateway 与执行引擎之间,负责:

  • 注册与健康管理引擎
  • 请求路由与调度(含 PD 分离场景)
  • 负载均衡与基础扩缩容钩子

依赖关系

PyPI 包名isagellm-control-plane导入命名空间sagellm_control

依赖(以 pyproject.toml 为准):

  • isagellm-protocol>=0.4.0.0,<0.5.0
  • pydantic>=2.0.0
  • httpx>=0.24.0

可选(用于本地直接执行引擎):

  • isagellm-core(提供 LLMEngine

被以下组件使用(示例):

  • sagellm(统一入口/CLI)
  • sagellm-gateway(API 网关)

安装指南

pip install isagellm-control-plane

Requirements: Python 3.11+

快速开始(CPU-first,可运行)

import asyncio

from sagellm_control import ControlPlaneManager, EngineState, ExecutionInstanceType
from sagellm_protocol import Request


async def main() -> None:
    cp = ControlPlaneManager(scheduling_policy="fifo", routing_strategy="least_loaded", mode="local")

    cp.register_engine(
        engine_id="engine-001",
        model_id="Qwen2-7B",
        host="localhost",
        port=8001,
        engine_kind="llm",
        metadata={"instance_type": ExecutionInstanceType.GENERAL.value},
    )
    cp.update_engine_state("engine-001", EngineState.READY)

    req = Request(
        request_id="req-001",
        trace_id="trace-001",
        model="Qwen2-7B",
        prompt="Hello",
        max_tokens=16,
        stream=False,
    )

    decision = await cp.schedule_request(
        request_id=req.request_id,
        trace_id=req.trace_id,
        model_id=req.model,
        prompt=req.prompt,
        max_tokens=req.max_tokens,
    )
    print(decision)

    cp.unregister_engine("engine-001")


if __name__ == "__main__":
    asyncio.run(main())

完整演示见 examples/mvp_integration_demo.py

Scheduler IR 模块使用说明

Control Plane 内置 Scheduler IR(Intermediate Representation)模块,用于把请求调度过程表达为可优化图结构:

  • IRBuilder:把请求构建为 SchedulerIR(Task/Prefill/Decode 节点 + 依赖边)
  • IROptimizer:执行可插拔优化 Pass(如 KVReusePassComputeCommOverlapPass
  • DefaultIRExecutor:把 IR 翻译为执行命令并通过 Control Plane/Engine Client 执行

可直接从根包导入:

from sagellm_control import IRBuilder, IROptimizer, DefaultIRExecutor

示例程序(至少 3 个):

运行示例:

python examples/ir_basic_example.py
python examples/ir_optimization_example.py
python examples/ir_kv_aware_example.py

API 文档(核心接口)

  • ControlPlaneManager
    • register_engine() / unregister_engine() / list_engines()
    • schedule_request()
    • execute_request() / stream_request()
    • get_embeddings()
  • EngineClient(HTTP 调用执行引擎)
  • LocalEngineClient(本地直接调用引擎)
  • Scheduler IRSchedulerIR, IRBuilder, IROptimizer, DefaultIRExecutor
  • SchedulingPolicy 及内置策略(FIFOPolicy, PriorityPolicy, SLOAwarePolicy, AdaptivePolicy, KVAwareSchedulingPolicy
  • 关键类型:EngineInfo, EngineState, SchedulingDecision, RequestPriority, RequestType

架构图示

flowchart LR
    A[Gateway/Client] --> B[ControlPlaneManager]
    B --> C[SchedulingPolicy]
    B --> D[RequestRouter/LoadBalancer]
    B --> E[EngineLifecycleManager]
    D --> F[EngineClient (HTTP)]
    B --> G[LocalEngineClient]
    F --> H[Execution Engines]
    G --> H

代码结构

sagellm_control/
├── types.py             # Core data types (EngineInfo, SchedulingDecision, etc.)
├── policies/            # Scheduling policies (FIFO, Priority, SLO-aware, Adaptive)
├── router.py            # Request routing and load balancing
├── lifecycle.py         # Engine lifecycle management
├── scaling.py           # Scaling manager (MVP hooks)
├── engine_client.py     # HTTP client to engines
├── local_engine_client.py # Local (in-process) engine client
├── ir/                  # Scheduler IR (types/builder/optimizer/executor)
└── manager.py           # ControlPlaneManager

开发指南

git clone git@github.com:intellistream/sagellm-control-plane.git
cd sagellm-control-plane
./quickstart.sh --dev

Quickstart 模式说明:

  • --standard:依赖优先从 PyPI 安装,当前仓库本地 editable 安装(稳定/发布导向)
  • --dev:在 standard 基础上,自动尝试将本地相邻子仓库切换为 editable(--no-deps 覆盖)
  • 每次安装前会动态清理已安装的 isagellm-* 历史包,保证流程可重入
# 标准模式(默认稳定依赖)
./quickstart.sh --standard

# 开发联调模式(本地 editable 覆盖)
./quickstart.sh --dev

# 查看帮助
./quickstart.sh --help

# 或手动安装
pip install -e ".[dev]"

运行测试:

pytest tests/ -v

Lint/格式化:

ruff format .
ruff check . --fix

提交流程:

  • 创建 Issue
  • fix/#123-xxx 分支开发
  • 提交 PR 到 main-dev

版本信息

Related Repositories


License

Proprietary - IntelliStream

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

isagellm_control_plane-0.5.4.9.tar.gz (172.2 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

isagellm_control_plane-0.5.4.9-py2.py3-none-any.whl (208.1 kB view details)

Uploaded Python 2Python 3

File details

Details for the file isagellm_control_plane-0.5.4.9.tar.gz.

File metadata

  • Download URL: isagellm_control_plane-0.5.4.9.tar.gz
  • Upload date:
  • Size: 172.2 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.14

File hashes

Hashes for isagellm_control_plane-0.5.4.9.tar.gz
Algorithm Hash digest
SHA256 d6079e558a12d88083bccd891533eb8bf3e0257e8f6b73397cfe4e9f5a184d49
MD5 f5d6b9e6b43ff22782e0ad3a553aba2f
BLAKE2b-256 9b810b5caf4c6e6c0ab840b406e589b677775a0317e221e1c3649283bfdc1a76

See more details on using hashes here.

File details

Details for the file isagellm_control_plane-0.5.4.9-py2.py3-none-any.whl.

File metadata

File hashes

Hashes for isagellm_control_plane-0.5.4.9-py2.py3-none-any.whl
Algorithm Hash digest
SHA256 9644adde62a69208c632589f4ee00aa5ca235de1bf6ebab069a06907744c58ef
MD5 c1b39f00ea35b44772b6bff774b59493
BLAKE2b-256 77f73360a65a5f29d81b41a03f12f8cf9f246f820f974ea25fac9be09b7f8ac9

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page