How aws/deep-learning-containers works
AWS Deep Learning Containers are prebuilt AI/ML container images for AWS workloads. The product spans framework and accelerator image variants, with SageMaker serving adapters and specialized inference services providing user-facing workflows. The clearest sampled paths are SageMaker request handling for OpenFold3 and TensorFlow, audio transcription with WhisperX, and Lambda engine startup. The tree and README establish broader image families, but unsampled files do not support detailed runtime links for most of them; those branches are represented without inferred call chains.
Main entry points: docker Framework Images · python_service.py TF HTTP Routes · sagemaker_serve.py OpenFold HTTP Adapter · server.py WhisperX API
Components 19
Container Images
- Framework Images
docker - CUDA Base Images
docker/base - PyTorch Images
docker/pytorch - TensorFlow Images
docker/tensorflow - Other Workloads
docker
SageMaker Serving
- TF Service Manager
scripts/docker/tensorflow/inference/sagemaker/serve.py - TF HTTP Routes
scripts/docker/tensorflow/inference/sagemaker/python_service.py - TF Serving Utilities
scripts/docker/tensorflow/inference/sagemaker/tfs_utils.py - Multi-Model Lock
scripts/docker/tensorflow/inference/sagemaker/multi_model_utils.py - OpenFold HTTP Adapter
scripts/docker/openfold3/sagemaker_serve.py - OpenFold Prediction
scripts/docker/openfold3/sagemaker_handler.py - AutoGluon Adapter
scripts/docker/autogluon/server.py
Inference Services
- WhisperX API
scripts/docker/whisperx/server.py - Transcription Model
scripts/docker/whisperx/server.py
Lambda Serving
- Lambda Runtime
- Engine Extension
scripts/docker/lambda/engine_extension.py - Invocation Handler
scripts/docker/lambda - vLLM or SGLang
Other components
- AWS Workload User
Connections 17
- AWS Workload User → Framework Images: uses images
- AWS Workload User → TF HTTP Routes: invokes
scripts/docker/tensorflow/inference/sagemaker/python_service.py - TF Service Manager → TF Serving Utilities: uses utilities
scripts/docker/tensorflow/inference/sagemaker/serve.py - TF HTTP Routes → TF Serving Utilities: parses requests
scripts/docker/tensorflow/inference/sagemaker/python_service.py - TF HTTP Routes → Multi-Model Lock: imports lock
scripts/docker/tensorflow/inference/sagemaker/python_service.py - AWS Workload User → OpenFold HTTP Adapter: invokes
scripts/docker/openfold3/sagemaker_serve.py - OpenFold HTTP Adapter → OpenFold Prediction: calls handler
scripts/docker/openfold3/sagemaker_serve.py - AWS Workload User → AutoGluon Adapter: invokes
scripts/docker/autogluon/server.py - AWS Workload User → WhisperX API: sends audio
scripts/docker/whisperx/server.py - WhisperX API → Transcription Model: transcribes
scripts/docker/whisperx/server.py - Lambda Runtime → Engine Extension: starts extension
scripts/docker/lambda/engine_extension.py - Engine Extension → vLLM or SGLang: starts server
scripts/docker/lambda/engine_extension.py - Lambda Runtime → Invocation Handler: dispatches requests
- CUDA Base Images → Framework Images: supports variants
- PyTorch Images → Framework Images: provides variants
- TensorFlow Images → Framework Images: provides variants
- Other Workloads → Framework Images: provides variants