--- title: Modal Clusters are generally available | daily.dev description: Modal Clusters, a multi-node GPU scheduling primitive, is now generally available via a single @modal.clustered decorator. Clusters support RDMA networking up... canonical: https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Modal Clusters are generally available | daily.dev og:description: Modal Clusters, a multi-node GPU scheduling primitive, is now generally available via a single @modal.clustered decorator. Clusters support RDMA networking up... og:url: https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy og:image: https://api.daily.dev/og/posts/xRXTKg1vy.png og:image:alt: Modal Clusters are generally available 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. # Modal Clusters are generally available **[Modal](https://daily.dev/sources/modal_labs)** · 7 min read · 2 upvotes · 0 comments ## Summary Modal Clusters, a multi-node GPU scheduling primitive, is now generally available via a single @modal.clustered decorator. Clusters support RDMA networking up to 6.4 Tbps, gang scheduling from Modal's shared capacity pool, and per-second billing instead of hourly or reservation-based pricing. Modal built a new gang scheduler with an observe-plan-act loop and added RDMA support to the gVisor container runtime, upstreaming the changes. Customers including Decagon, 1x, and Runway describe using Clusters for large-scale model fine-tuning, robotics world-model pretraining, and multi-node video generation inference. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How do I create a multi-node GPU cluster on Modal with RDMA enabled? Use the @modal.clustered(size=N, rdma=True) decorator on top of an @app.function with a gpu argument like gpu="B300:8". This automatically configures RDMA for PyTorch and NCCL across nodes, exposes container IPs and rank via modal.Cluster.from_context(), and bills by the second once the cluster is scheduled from Modal's shared capacity pool. _daily.dev surfaces practical setup details like this for engineers configuring multi-node GPU training._ ### How fast is RDMA networking compared to TCP for syncing model weights during training? RDMA can push up to 6.4 Tbps versus roughly 50 Gbps over standard TCP. For example, syncing about 717 GB of BF16 weights from a trainer to a rollout engine during a GLM 4.7 training step takes nearly two minutes over 50 Gbps TCP but under two seconds over RDMA, repeated across thousands of steps. _Developers comparing networking options for distributed training can track benchmarks like this on daily.dev._ ### Why is RDMA difficult to support across different cloud providers? Each cloud provider implements RDMA differently, requiring distinct drivers, environment variables, and userspace libraries, which makes a unified integration hard to build. Modal addressed this by automating RDMA setup across clouds and adding RDMA support directly into the gVisor container runtime, upstreaming the changes to the open-source project. _daily.dev helps engineers wrangling cross-cloud infrastructure quirks like RDMA stay on top of fixes._ ## Similar posts on daily.dev - [Product Updates: Modal in AWS & GCP marketplaces, Sandbox improvements, and more](https://daily.dev/posts/product-updates-modal-in-aws-gcp-marketplaces-sandbox-improvements-and-more-xdksa0e6m) · Modal · 1 upvotes · 0 comments - [Modal \+ Datalab: Deploy high-throughput document intelligence in <5 minutes](https://daily.dev/posts/modal-datalab-deploy-high-throughput-document-intelligence-in-5-minutes-arlmayzdr) · Modal · 0 upvotes · 0 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#gpu](https://daily.dev/tags/gpu) [View this post on daily.dev](https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy) ```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":"Modal Clusters are generally available","url":"https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy"},"datePublished":"2026-10-01T07:16:35.189Z","dateModified":"2026-10-01T07:16:56.907Z","description":"Modal Clusters, a multi-node GPU scheduling primitive, is now generally available via a single @modal.clustered decorator. Clusters support RDMA networking up...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6b8ecee9973d88e4d92f78059b9f36cc?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/6b8ecee9973d88e4d92f78059b9f36cc?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Modal","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":"Modal","logo":"https://media.daily.dev/image/upload/s--HfKZHzC1--/f_auto/v1750946306/logos/modal_labs","url":"https://daily.dev/sources/modal_labs"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,gpu","timeRequired":"PT7M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Modal","item":"https://daily.dev/sources/modal_labs"},{"@type":"ListItem","position":3,"name":"Modal Clusters are generally available"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/modal-clusters-are-generally-available-xrxtkg1vy#faq","mainEntity":[{"@type":"Question","name":"How do I create a multi-node GPU cluster on Modal with RDMA enabled?","acceptedAnswer":{"@type":"Answer","text":"Use the @modal.clustered(size=N, rdma=True) decorator on top of an @app.function with a gpu argument like gpu=\"B300:8\". This automatically configures RDMA for PyTorch and NCCL across nodes, exposes container IPs and rank via modal.Cluster.from_context(), and bills by the second once the cluster is scheduled from Modal's shared capacity pool. daily.dev surfaces practical setup details like this for engineers configuring multi-node GPU training."}},{"@type":"Question","name":"How fast is RDMA networking compared to TCP for syncing model weights during training?","acceptedAnswer":{"@type":"Answer","text":"RDMA can push up to 6.4 Tbps versus roughly 50 Gbps over standard TCP. For example, syncing about 717 GB of BF16 weights from a trainer to a rollout engine during a GLM 4.7 training step takes nearly two minutes over 50 Gbps TCP but under two seconds over RDMA, repeated across thousands of steps. Developers comparing networking options for distributed training can track benchmarks like this on daily.dev."}},{"@type":"Question","name":"Why is RDMA difficult to support across different cloud providers?","acceptedAnswer":{"@type":"Answer","text":"Each cloud provider implements RDMA differently, requiring distinct drivers, environment variables, and userspace libraries, which makes a unified integration hard to build. Modal addressed this by automating RDMA setup across clouds and adding RDMA support directly into the gVisor container runtime, upstreaming the changes to the open-source project. daily.dev helps engineers wrangling cross-cloud infrastructure quirks like RDMA stay on top of fixes."}}]} ```