A reproducible path from model packaging and inference serving to canaries, observability, and rollback.
Model deployment engineering
A reproducible path from model packaging and inference serving to canaries, observability, and rollback.
7 chapters3 casesPublished
LEARNING PATH
Tutorial chapters
Actionable chapters arranged in a useful learning sequence.
chapters · Published模型部署基础:可重复的模型打包
固定模型、Tokenizer、依赖和配置,生成可审计的部署工件。
Read chapter chapters · Published推理服务接口:健康检查与兼容 API用稳定的 HTTP 契约隔离模型运行时,明确超时、错误和请求上限。
Read chapter chapters · PublishedvLLM 服务化:从本地兼容接口开始使用官方 OpenAI 兼容接口启动本地模型,并验证鉴权和请求边界。
Read chapter chapters · Published推理性能与容量规划:先测边界再调参数用可复现的输入集测量延迟、并发、上下文和显存边界。
Read chapter chapters · PublishedKubernetes 发布:探针、资源与滚动更新为模型服务配置资源边界、探针和可回滚的 Deployment。
Read chapter chapters · Published模型服务观测与回滚:让故障可定位记录请求级指标、模型版本和错误类别,在异常时快速切回上一版。
Read chapter chapters · Published模型量化与质量预算:先定义可接受损失在推理成本、延迟和回答质量之间建立可测量的量化发布门槛。
Read chapterPRACTICE
Practical cases
Reusable notes built around tools, workflows, and real tasks.
