RIDE is the inference Control Plane. Request Throttling is one Controller type.
Ray is the process execution substrate: RIDE deploys independent Controller and
vllm serve jobs, observes them, and manages only its persisted owned jobs.
RIDE connects to an existing RayCluster. It never requires a new Coder template
or workspace per deployment. An existing owner-only Coder App is optional. RIDE's management API is not on the model request path.
先从 已有 Ray 集群的命令行操作说明 开始。
RIDE × Request Throttling 一页总览 集中展示系统责任边界、官方迁移、三方真机性能、全程图、原始数据索引、 一键回收/报告及未完成项。RIDE 是控制面,Controller 的性能结果单独验收; CPU/CI 和 payload 检查不冒充实际 NPU 端到端证据。
新增 Controller 见 Controller 开发与接入指南。
安装部署见 复用已有 RayCluster 的安装与部署脚本。
Current deployment authority: existing-Ray deployment policy. CPU tests are prerequisites; complete acceptance requires real NPU/model/control actions.
python -m pip install '.[ray]'
ride start --ray-address http://127.0.0.1:8265
ride start model-service vllm --name model-main \
--model /models/Qwen --endpoint http://model-host:8000 \
--port 8000 --num-gpus 1
ride start controller request-throttling --name throttle-main \
--backend http://model-host:8000 --endpoint http://controller-host:9000 \
--port 9000 --executor ray
ride get model-services
ride get controllers
ride describe controller throttle-main
ride logs model-service model-main
ride stop controller throttle-main
ride start controller throttle-mainThe endpoints must be reachable from the management process and clients;
localhost does not address another Ray node. GPU reservation uses
--num-gpus; Ascend custom resources use e.g.
--resources '{"NPU": 1}', according to the actual Ray resource names.
CPU/GPU requirements and --runtime-env must match the deployed environment.
Model weights and pinned Controller dependencies must already exist in that
runtime. RIDE does not download weights or assume accelerator compatibility.
Names identify instances; request-throttling and vllm identify types.
Same-name same-configuration start is idempotent; different configuration is
rejected. YAML is not required. Optional --config currently reads JSON;
YAML parsing is not implemented. ride serve --config examples/ride.json
remains compatible. ride start runs the management process in the foreground.
Clients use --address or RIDE_ADDRESS to reach it.
Use --model-service model-main on Controller creation to require an owned
model service to be Ready and its endpoint to match the Controller backend.
Without that option, an external model backend is allowed. Stopping a
Controller leaves both owned and external model services running.
GET /healthz: management process health.GET /api/v1/status: workspace identity, independent model and Controller states.GET /api/v1/{controllers|model-services}: list instances.GET /api/v1/{kind}/{name}: inspect an instance.GET /api/v1/{kind}/{name}/logs: owned Ray job logs.POST /api/v1/{kind}/{name}/start: create/start; JSON spec optional for existing instance.POST /api/v1/{kind}/{name}/stop: stop only that persisted owned job.
UI and CLI use the same API. Ray Running and service Ready are distinct; neither
proves effective control or performance improvement. Submission outcomes that
are uncertain retain the exact Job ID; a failed lookup does not create another
Job. Persistent .ride/state.json supports management restart/adoption.
Ray execution only; a local executor, generic third-party discovery, full common control-loop contracts, graceful drain, high availability, authentication inside the listener, and cross-Controller arbitration are not implemented. Restrict the listener to the workspace and use owner-only Coder access. A trusted owner can provide runtime environments and model arguments. Stop can interrupt in-flight requests; it is not graceful drain.
The unvalidated extension-manager Bundle draft is excluded from the v0.2 wheel. See design, refactor acceptance and historical workspace acceptance.
Standardized acceptance is built in: ride verify --controller <instance> --model-service <instance> --output evidence.json. Add --lifecycle to explicitly test stop/restore, --model <served-name> for a real inference, and --baseline <before-restart.json> to verify management restart adoption. See acceptance contract. PASS only covers executed checks; NOT_RUN and performance scope remain explicit.
Coding agents and reviewers: follow mandatory acceptance procedure, using the plan and action templates in acceptance/.