AI Trending
All 全部 Deep Analysis 深度解析 Foresight 前瞻 AI News AI资讯 Open Source 开源项目 AI Products AI产品 Research Papers 论文研究 AI Security AI安全 AI Practices AI实践 AI Skills AI技能 AI Overseas AI出海
AI Practices AI实践 4h ago • Updated 1h ago 更新于 1小时前 43

Bring your own model with Amazon SageMaker AI: Script mode in SDK v3 使用 Amazon SageMaker AI 自带模型：SDK v3 中的脚本模式

Amazon SageMaker Python SDK v3 introduces a unified `ModelTrainer` and `ModelBuilder` API, replacing the fragmented v2 estimator classes (SKLearn, PyTorch, XGBoost) with a single interface for all frameworks The new `SourceCode` configuration object syncs local source directories into training containers at runtime, eliminating the need to rebuild Docker images for code changes Users can bring any container image (custom-built, AWS Deep Learning Containers, or third-party) while the SDK handles Amazon SageMaker AI SDK v3 引入全新架构，用统一的 ModelTrainer 和 ModelBuilder 替代 v2 中按框架区分的 Estimator 类族 新增 SourceCode 配置对象，支持在训练/推理任务运行时动态同步本地代码目录，无需将代码打包进容器镜像 支持任意容器来源（自建镜像、AWS DLC 或第三方镜像），实现更快的迭代速度和更完整的容器控制 提供两个端到端示例：基于 scikit-learn 的表格 ML 工作流（使用 DJL Serving 部署）和基于 Stable Diffusion 3.5 + LoRA 的多 GPU 分布式微调

60
Hot 热度
65
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • Amazon SageMaker Python SDK v3 introduces a unified ModelTrainer and ModelBuilder API, replacing the fragmented v2 estimator classes (SKLearn, PyTorch, XGBoost) with a single interface for all frameworks
  • The new SourceCode configuration object syncs local source directories into training containers at runtime, eliminating the need to rebuild Docker images for code changes
  • Users can bring any container image (custom-built, AWS Deep Learning Containers, or third-party) while the SDK handles code injection automatically
  • Two end-to-end examples demonstrate the workflow: a scikit-learn Random Forest on the diabetes dataset deployed via DJL Serving, and fine-tuning Stable Diffusion 3.5 with LoRA using multi-GPU distributed training

Why It Matters

The SDK v3 redesign significantly lowers the barrier for practitioners who want to run custom ML workflows on SageMaker without managing complex Docker image pipelines. By unifying the API across frameworks and enabling runtime code injection, it accelerates iteration cycles and reduces operational overhead for both traditional ML and generative AI workloads.

Technical Details

  • Unified API Architecture: ModelTrainer replaces all v2 framework-specific estimators for training jobs, while ModelBuilder replaces the Model/Predictor pattern for deployment, with predictions handled through invoke()
  • SourceCode Object: Accepts a source_dir path plus either a command string (training) or entry_script (inference); at job launch, SageMaker syncs the directory into the container without baking code into the image
  • Container Flexibility: Supports any ECR-hosted image—user-built, AWS Deep Learning Containers, or third-party—giving full control over system packages, CUDA libraries, and runtime dependencies
  • Dependency Management: requirements.txt placed in source_dir is still supported for Python dependency installation alongside the container's base environment
  • Example Workflows: The scikit-learn example uses a minimal python:3.13-slim base image with build-essential, jq, and git; the generative AI example leverages Hugging Face Accelerate for multi-GPU distributed training of Stable Diffusion 3.5 with LoRA

Industry Insight

  • The shift toward a unified, framework-agnostic API reflects the industry trend of abstracting away infrastructure complexity so practitioners can focus on model development rather than container orchestration—expect similar unification patterns across cloud ML platforms
  • Runtime code injection via SourceCode enables a true CI/CD-friendly workflow where data scientists can iterate on training logic at the same speed as local development, reducing the friction between experimentation and production deployment
  • Support for bringing custom containers alongside managed images bridges the gap between flexibility and convenience, making SageMaker more competitive against platforms like Vertex AI and Azure ML for teams with specialized inference requirements (e.g., custom C++ binaries or DJL Serving)

TL;DR

  • Amazon SageMaker AI SDK v3 引入全新架构，用统一的 ModelTrainer 和 ModelBuilder 替代 v2 中按框架区分的 Estimator 类族
  • 新增 SourceCode 配置对象，支持在训练/推理任务运行时动态同步本地代码目录，无需将代码打包进容器镜像
  • 支持任意容器来源（自建镜像、AWS DLC 或第三方镜像），实现更快的迭代速度和更完整的容器控制
  • 提供两个端到端示例：基于 scikit-learn 的表格 ML 工作流（使用 DJL Serving 部署）和基于 Stable Diffusion 3.5 + LoRA 的多 GPU 分布式微调

为什么值得看

本文展示了 SageMaker Python SDK v3 的核心架构变革，为 AI 从业者提供了更灵活、更高效的自定义模型训练与部署范式，尤其适合需要频繁迭代算法或混合使用多种框架的团队。

技术解析

  • 统一 API 设计：v3 用单个 ModelTrainer 类替代 v2 的 SKLearn/PyTorch/XGBoost 等框架专属 Estimator，用 ModelBuilder 替代 Model + Predictor 模式，实现跨框架一致接口。
  • SourceCode 运行时注入：通过 SourceCode(source_dir=..., command=...) 或 entry_script=... 配置，SDK 在作业启动时将本地代码目录同步进容器，代码无需 bake 进镜像，支持快速迭代。
  • 容器灵活性：支持任意 ECR 镜像（自建、AWS DLC 或第三方），用户可自由安装系统包、CUDA 库等，SDK 不预设容器内容。
  • 依赖管理：继续使用 requirements.txt 管理 Python 依赖，但框架运行时环境由用户控制的容器镜像决定。
  • 示例工作流：示例一展示 scikit-learn Random Forest 训练与 DJL Serving 部署；示例二展示 Stable Diffusion 3.5 + LoRA 多 GPU 分布式微调（使用 Hugging Face Accelerate）。

行业启示

  • MLOps 效率提升：代码与镜像分离的架构显著减少迭代周期，适合需要频繁实验的 AI 研发团队。
  • 框架无关性趋势：统一 API 降低多框架协作门槛，推动企业向标准化 ML 流水线演进。
  • 自定义容器价值凸显：SDK v3 鼓励用户构建可控的基础镜像，结合运行时代码注入，兼顾灵活性与可复现性。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明：以上内容由 AI 生成，仅供参考。

Deployment 部署 Training 训练 Inference 推理 Product Launch 产品发布
Read Original → 阅读原文 →

Related Articles 相关文章

[GitHub] keras-team/keras Keras 3 多后端深度学习框架
[GitHub] tensorflow/tensorflow [GitHub] TensorFlow项目
Why vLLM and SGLang Are Replacing Ollama for Agentic Workflows 为什么 vLLM 和 SGLang 正在取代 Ollama 用于智能体工作流
How GoDaddy transformed its analytics with Amazon QuickSight GoDaddy如何利用Amazon QuickSight转型其分析系统
[GitHub] lucidrains/vit-pytorch GitHub lucidrains/vit-pytorch 项目