PyTorch Developer Podcast

API design via lexical and dynamic scoping

Intro to distributed

Double backwards

Functional modules

CUDA graphs

Default arguments

Anatomy of a domain library

TensorAccessor

This podcast episode introduces TensorAccessor, a data structure in PyTorch that allows efficient element-by-element tensor manipulation, including for tensors with complex layouts, by providing direct access to strided data without contiguous memory allocation.

Random number generators

vmap

Expect tests

XLA

TH

TorchScript

CMake

torchdeploy

C++ frontend

PyObject preservation

Mobile selective build

torch.nn