The PyTorch Developer Podcast is a place for the PyTorch dev team to do bite sized (10-20 min) topics about all sorts of internal development topics in PyTorch.
This podcast episode delves into weak references, their implementation, and consequences in PyTorch and C++. It explores use cases and implementation details of weak references, discussing their role in caching and how they allow objects to be garbage collected even with remaining references. The episode compares C++'s...
This podcast episode thoroughly explores the concept of strides in PyTorch, delving into their importance in representing multidimensional tensor data, their role in creating views and performance optimization, and their limitations in handling non-contiguous inputs. It also discusses the concepts of strided tensors, m...
This podcast episode explores AOT AutoGrad, a revolutionary framework that seamlessly integrates compilers into PyTorch training, enabling substantial improvements in performance, memory usage, and code efficiency. By considering the operations during training as tensor operations similar to those in inference, AOT Aut...
This podcast episode delves into the inner workings of PyTorch's dispatcher, examining its history, significance, and various aspects such as dispatch keys, multi-dispatch, and backend selection. It also explores the evolution of the dispatcher and the motivations behind design decisions, showcasing how PyTorch continu...
This podcast episode explores the different APIs in the PyTorch library, namely the Torch API and the A10 API, and delves into the relationship between PyTorch and Aten. It further discusses the conversion processes between Torch and A10, the potential for a third API known as PrimOps, and the significance of PrimOps i...
Follow this podcast in Podwise
Sign in to get AI summaries, transcripts and mind maps for any episode, including new ones.
