What problem does it solve? Choosing the wrong SageMaker serving container, a stale tag, or an incompatible CUDA/AMI combination produces opaque deployment failures like "Failed to pass health check" that waste hours of debugging. This Skill removes guesswork by enforcing verified image selection from AWS's official Deep Learning Containers catalog instead of inferring URIs from memory. ## Core Features & Use Cases - Model-to-container decision rules: Maps model families (LLMs, multimodal, embeddings, rerankers, diffusion, classifiers) to the correct container family such as HuggingFace vLLM, vLLM-Omni, TEI, or DJL Inference, with HuggingFace-curated images mandatory when compatible. - Failure prevention: Documents the cu130 AMI requirement, TEI GPU/CPU variant selection, VPC/NAT image-pull constraints, and known-broken images like the NCCL-defective huggingface-pytorch-inference GPU tags. - Image mirroring script: Includes scripts/mirror_image.py to copy public ECR images into private ECR for VPC deployments without NAT gateways. - Use Case: Before deploying Qwen3-Reranker to SageMaker, preflight its config.json to discover it is a generative reranker, then select the HuggingFace vLLM DLC URI for your region instead of TEI, avoiding a 20-minute failed endpoint cycle. ## Quick Start Ask the assistant to select and verify the correct SageMaker serving container URI for your HuggingFace model and target AWS region before writing any deployment code.
