Features

The "features" lens groups the user-visible capabilities that span multiple subsystems. Where Systems is "what's in the box", this is "what can I do with it".

Page	What it covers
Eager execution	The default mode: imperative torch.* calls, dynamic autograd
torch.compile	Graph capture + compilation via Dynamo + AOT autograd + Inductor
torch.export and AOTInductor	Production graph capture and ahead-of-time compilation
Distributed training	DDP, FSDP, DTensor, pipelining, multi-node
Mixed precision and autocast	torch.amp, torch.autocast, fp16/bf16/fp8
Profiling and tracing	torch.profiler, memory snapshots, structured trace + tlparse
Quantization	Eager / FX / PT2E quantization workflows
Mobile and edge	PyTorch Mobile, ExecuTorch, on-device inference
Custom ops and extensions	torch.library, torch.utils.cpp_extension, custom autograd functions
Sparse and quantized tensors	Sparse layouts, nested tensors, masked tensors

The features cross-link heavily into the systems pages because almost every feature is implemented across multiple subsystems.