<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Pytorch on Sandeep Kumar</title><link>https://sandeep007734.github.io/tags/pytorch/</link><description>Recent content in Pytorch on Sandeep Kumar</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Wed, 12 Aug 2026 00:00:00 +0530</lastBuildDate><atom:link href="https://sandeep007734.github.io/tags/pytorch/feed.xml" rel="self" type="application/rss+xml"/><item><title>PyTorch Distributed Training Internals</title><link>https://sandeep007734.github.io/posts/pytorch_distributed_training/</link><pubDate>Wed, 12 Aug 2026 00:00:00 +0530</pubDate><guid>https://sandeep007734.github.io/posts/pytorch_distributed_training/</guid><description>How DistributedDataParallel differs from DataParallel, how NCCL implements the gradient all-reduce, how bucketing overlaps communication with the backward pass, and how to tell a dataloader stall from a compute bottleneck on a profiler timeline.</description></item></channel></rss>