An open-source deep learning framework (originally Meta, now the PyTorch Foundation) built on tensors, autograd, and dynamic computation graphs — the graph is defined as the code runs.
Why it dominates
Debuggable like ordinary Python, which made it the default for research; the ecosystem (Hugging Face Transformers, torchvision, Lightning) followed.
Where it runs on AWS
AWS Deep Learning AMIs and Containers — preinstalled with GPU drivers and CUDA.
Amazon SageMaker AI — managed training jobs and endpoints with a first-class PyTorch estimator.
On Trainium and Inferentia chips through the AWS Neuron SDK. See AWS ML Infrastructure.
Where it sits in the stack
Tier 3, ML Frameworks — maximum control, maximum operational responsibility.
Not to be confused with
TensorFlow — the other major framework, historically static-graph and production-oriented. Both are equally supported on AWS; exam questions never hinge on choosing between them.