--- title: A Ray-Focused Guide to PyTorch Conference North America... description: PyTorch Conference North America 2026 in San Jose is highlighting Ray as foundational infrastructure for scaling AI across the full lifecycle: data curation,... canonical: https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: A Ray-Focused Guide to PyTorch Conference North America – PyTorch | daily.dev og:description: PyTorch Conference North America 2026 in San Jose is highlighting Ray as foundational infrastructure for scaling AI across the full lifecycle: data curation,... og:url: https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws og:image: https://api.daily.dev/og/posts/nbKpHwiwS.png og:image:alt: A Ray-Focused Guide to PyTorch Conference North America – PyTorch og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # A Ray-Focused Guide to PyTorch Conference North America – PyTorch **[PyTorch](https://daily.dev/sources/pytorch)** · 8 min read · 0 upvotes · 0 comments ## Summary PyTorch Conference North America 2026 in San Jose is highlighting Ray as foundational infrastructure for scaling AI across the full lifecycle: data curation, distributed training, and production serving. Sessions include a keynote from Ray co-creator Ion Stoica on Ray and Kubernetes convergence, LinkedIn's three-layer Ray-based PyTorch training stack (cutting dataloader memory 50-70%), Uber Eats' migration from TensorFlow/Horovod to PyTorch with Ray Data (5x speedup, 90% lower memory), Pinterest's PTEnv post-training harness, and Anyscale's SkyRL for async RL training on 350B+ parameter MoE models. The piece frames Ray as the standard substrate for post-training and RL at labs like Cursor, Microsoft AI, and NVIDIA. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How did LinkedIn reduce memory usage in their Ray-based PyTorch training pipeline? LinkedIn built a three-layer Ray-based training stack that keeps data loading off the GPUs by routing it through dedicated CPU nodes over a zero-copy bridge, using a Ray-actor-based data loading library streaming Avro, Parquet, and Iceberg formats alongside a Ray Train layer with FSDP, HSDP, elastic scaling, and async checkpointing. This cut dataloader memory usage by 50-70%, though they had to work around Ray in places to reclaim 24% of step time. _daily.dev surfaces practical distributed-training engineering stories like this for teams scaling PyTorch._ ### What results did Uber get from migrating its Eats recommendation system from TensorFlow to PyTorch with Ray Data? Uber's migration of its Eats recommendation stack from TensorFlow/Horovod to PyTorch, using Ray Data to replace legacy Spark-based preprocessing with distributed feature-stat computation and zero-copy batch transformations, delivered a 5x speedup and 90% lower memory usage for data transformation plus a 20x improvement in training throughput. _Developers evaluating a TensorFlow-to-PyTorch migration can track real-world outcomes like this on daily.dev._ ### What is Ray Direct Transport and how does it help scale post-training workloads? Ray Direct Transport is a performance optimization in Ray Core that moves PyTorch tensors directly between GPUs, reducing overhead during post-training and reinforcement learning loops. It's part of continuous engineering work establishing Ray Core as scheduling infrastructure across clusters of up to 10,000 nodes, underpinning post-training frameworks such as SkyRL, Miles, and NeMo RL. _Teams building post-training infrastructure can follow Ray Core scalability developments on daily.dev._ ## Similar posts on daily.dev - [Parallel Python at Any Scale with Ray](https://daily.dev/posts/parallel-python-at-any-scale-with-ray-g47ol7ob4) · Planet Python · 0 upvotes · 0 comments - [Ray AI Engine Pulled Into The PyTorch Foundation For Unified Open AI Compute Stack](https://daily.dev/posts/ray-ai-engine-pulled-into-the-pytorch-foundation-for-unified-open-ai-compute-stack-2drk5shft) · Phoronix · 0 upvotes · 0 comments --- Tags: [#kubernetes](https://daily.dev/tags/kubernetes), [#distributed-systems](https://daily.dev/tags/distributed-systems), [#pytorch](https://daily.dev/tags/pytorch), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"A Ray-Focused Guide to PyTorch Conference North America – PyTorch","url":"https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws"},"datePublished":"2026-09-30T16:48:01.205Z","dateModified":"2026-09-30T16:48:26.807Z","description":"PyTorch Conference North America 2026 in San Jose is highlighting Ray as foundational infrastructure for scaling AI across the full lifecycle: data curation,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6c6e8ee318693c4de6f5f85f45160b68?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6c6e8ee318693c4de6f5f85f45160b68?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"PyTorch","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"PyTorch","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/5fdab3d75a834a97a52d433d9c7e5ff9","url":"https://daily.dev/sources/pytorch"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"kubernetes,distributed-systems,pytorch,reinforcement-learning","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"PyTorch","item":"https://daily.dev/sources/pytorch"},{"@type":"ListItem","position":3,"name":"A Ray-Focused Guide to PyTorch Conference North America – PyTorch"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/a-ray-focused-guide-to-pytorch-conference-north-america-pytorch-nbkphwiws#faq","mainEntity":[{"@type":"Question","name":"How did LinkedIn reduce memory usage in their Ray-based PyTorch training pipeline?","acceptedAnswer":{"@type":"Answer","text":"LinkedIn built a three-layer Ray-based training stack that keeps data loading off the GPUs by routing it through dedicated CPU nodes over a zero-copy bridge, using a Ray-actor-based data loading library streaming Avro, Parquet, and Iceberg formats alongside a Ray Train layer with FSDP, HSDP, elastic scaling, and async checkpointing. This cut dataloader memory usage by 50-70%, though they had to work around Ray in places to reclaim 24% of step time. daily.dev surfaces practical distributed-training engineering stories like this for teams scaling PyTorch."}},{"@type":"Question","name":"What results did Uber get from migrating its Eats recommendation system from TensorFlow to PyTorch with Ray Data?","acceptedAnswer":{"@type":"Answer","text":"Uber's migration of its Eats recommendation stack from TensorFlow/Horovod to PyTorch, using Ray Data to replace legacy Spark-based preprocessing with distributed feature-stat computation and zero-copy batch transformations, delivered a 5x speedup and 90% lower memory usage for data transformation plus a 20x improvement in training throughput. Developers evaluating a TensorFlow-to-PyTorch migration can track real-world outcomes like this on daily.dev."}},{"@type":"Question","name":"What is Ray Direct Transport and how does it help scale post-training workloads?","acceptedAnswer":{"@type":"Answer","text":"Ray Direct Transport is a performance optimization in Ray Core that moves PyTorch tensors directly between GPUs, reducing overhead during post-training and reinforcement learning loops. It's part of continuous engineering work establishing Ray Core as scheduling infrastructure across clusters of up to 10,000 nodes, underpinning post-training frameworks such as SkyRL, Miles, and NeMo RL. Teams building post-training infrastructure can follow Ray Core scalability developments on daily.dev."}}]} ```