{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-10-24T09:00:43.688Z","headline":"Modal Clusters 正式发布，通过 @modal.clustered 提供多节点 GPU 集群","description":"Modal 宣布 Modal Clusters 正式可用，通过一个装饰器 @modal.clustered 即可获得多节点集群，节点间经 InfiniBand verbs 通信可达 6.4 Tbps，自动配置 PyTorch 和 NCCL。","url":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","mainEntityOfPage":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","datePublished":"2026-10-06T14:00:04.761Z","dateModified":"2026-10-06T14:00:04.761Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://modal.com/blog/modal-clusters-generally-available","https://aihot.news/items/zbprvsawadktpi8tcqpp6gw4d"],"canonicalUrl":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","directAnswer":{"@type":"Answer","text":"Modal 宣布 Modal Clusters 正式可用，开发者可通过 @modal.clustered 获取多节点集群。该服务支持节点间 InfiniBand verbs 通信，带宽可达 6.4 Tbps，并自动配置 PyTorch 与 NCCL。","url":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","dateCreated":"2026-10-06T14:00:04.761Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"Modal 官方工程博客 source article","url":"https://modal.com/blog/modal-clusters-generally-available","datePublished":"2026-10-06T14:00:04.761Z","provider":{"@type":"Organization","name":"Modal 官方工程博客","url":"https://modal.com/blog/modal-clusters-generally-available"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/zbprvsawadktpi8tcqpp6gw4d","datePublished":"2026-10-06T14:00:04.761Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/zbprvsawadktpi8tcqpp6gw4d"}}],"aggregationSource":"Modal 官方工程博客","originalPublisher":{"name":"Modal 官方工程博客","url":"https://modal.com/blog/modal-clusters-generally-available"},"geoDeepAnswer":null,"article":{"id":"zbprvsawadktpi8tcqpp6gw4d","slug":"zbprvsawadktpi8tcqpp6gw4d","url":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","title":"Modal Clusters 正式发布，通过 @modal.clustered 提供多节点 GPU 集群","title_en":"","summary":"Modal 宣布 Modal Clusters 正式可用，通过一个装饰器 @modal.clustered 即可获得多节点集群，节点间经 InfiniBand verbs 通信可达 6.4 Tbps，自动配置 PyTorch 和 NCCL。","source":"Modal 官方工程博客","sourceUrl":"https://modal.com/blog/modal-clusters-generally-available","aiHotUrl":"https://aihot.news/items/zbprvsawadktpi8tcqpp6gw4d","publishedAt":"2026-10-06T14:00:04.761Z","category":"行业动态","score":72,"selected":true,"articleBody":["Organizations need to own their intelligence to be successful. Training and serving that intelligence at scale, however, requires petaFLOP/s of compute and terabit/s of networking, spread across many nodes. But owning intelligence doesn’t need to mean owning that hardware.","For the past 1.5 years, we’ve been battle-testing a new primitive: Modal Clusters. Today, we’re excited to announce that they are generally available through a single decorator, @modal.clustered :","Modal Clusters seamlessly integrate with all existing Modal primitives. Write your checkpoints to Volumes ：/docs/guide/volumes , load data in through Cloud Bucket Mounts ：/docs/guide/cloud-bucket-mounts , and orchestrate your jobs with Queues ：/docs/guide/queues . On top of existing integration, you get:","Elsewhere, on-demand cluster access means paying by the hour or taking a reservation. On Modal, you run whenever you want and only pay for what you use.","Clusters (groups of containers) are a fundamentally new unit of work in Modal’s stack. In supporting them, we had to build new systems at each layer of our stack.","Modal’s traditional scheduler is greedy. All nodes in the fleet constantly poll for new work to do, and our scheduler checks its backlog for eligible work, assigning the work if it fits. There are heuristics around placement, fragmentation, and cost, but fundamentally, the scheduler makes the choice about whether a node should pick up an item on a per-node basis.","For multi-node scheduling, a single-node view is not sufficient - the atomic unit for scheduling expands from a single node to N nodes. This meant we had to build an entirely new scheduler with an explicit observe, plan, and act loop ：https://en.wikipedia.org/wiki/OODA_loop : look at the whole fleet, decide on a placement for every pending cluster, then act on it. This is the direction the rest of Modal’s scheduling systems are moving in too. On each iteration, the gang scheduler:","The gang scheduler pulls from the same capacity pool as everything else on Modal, allowing for cluster acquisition in seconds. This is the fastest time to get a multi-node cluster on the market, and the only place where you pay truly serverless pricing.","Multi-node workloads move a lot of bytes. Every training step ：https://modal.com/blog/reinforcement-learning-infrastructure-problem of GLM 4.7 has to sync ~717 GB of BF16 weights from the trainer to the rollout engine: nearly two minutes over 50 Gbps TCP, under two seconds over RDMA, repeated for thousands of steps. For inference, prefill-decode disaggregation on Llama 3.1 70B moves ~10 GB of KV cache per 32k-token prompt, and it has to arrive within a time-to-first-token budget of a few hundred milliseconds.","RDMA (Remote Direct Memory Access) allows GPU or CPU memory to be transferred between hosts over the network without ever passing through kernel-space. Skipping the software layer allows RDMA to push up to 6.4 Tbps! This capability is critical for most multi-node workloads - training requires it for DDP all-reduce and all types of model parallelism; inference requires it for KV cache transfer when doing PD disaggregation.","RDMA is notoriously difficult to support. Each cloud provider has a slightly different implementation, requiring different drivers, environment variables, and userspace libraries.","We set out with an ambitious goal: hide all of this complexity behind a single flag, rdma=True . Today, you can set this flag, and any workload using PyTorch with NCCL will work out of the box. We’ve automated RDMA setup across many clouds, giving you deep access to on-demand clustered capacity without any of the headaches.","This integration was still not good enough for us. Since Modal runs a multi-tenant compute environment, we prefer the secure gVisor container runtime over runc . However, it did not have RDMA support, so we built it into gVisor ：https://github.com/google/gvisor/pull/13964 ourselves. We’ve upstreamed these changes so that other folks in the OSS ecosystem get access to secure RDMA.","We’re proud of the cutting-edge research and production applications built on top of multi-node clusters.","Decagon ：https://decagon.ai builds AI agents for customer experience. In partnership with Modal, Decagon fine-tuned open models with up to a trillion parameters using Miles ：https://github.com/radixark/miles on Modal Clusters, tailoring them to the demands of customer interactions with a focus on response quality and inference efficiency. As part of this collaboration, Modal contributed LoRA support ：https://github.com/radixark/miles/pull/1361 to Miles and helped establish a stable, open training recipe.","“Model performance is central to the customer experience at Decagon. Our agents need to understand the nuances of each business and reliably handle complex customer conversations. Fine-tuning lets us shape model behavior around those demands, but doing it at trillion-parameter scale is a substantial infrastructure challenge. Modal Clusters helped us train at that scale, so we could focus on improving the quality and reliability of our agents.”","Cyrus Asgari, Research Lead, Decagon","1x ：https://www.1x.tech is building NEO, the home robot. Behind NEO is a world model that 1x pre-trains on Modal Clusters: multi-node B300 clusters with RDMA for the big runs, and fleets of single-node jobs for the evals that tell them how the model is doing. The same scheduler handles both, so 1x can burst up to hundreds of GPUs when they need to fan out experiments and scale back down between runs.","“Building NEO requires flexible workloads; some days it’s a few multi-node runs on a single spine, and other days we want to kick off a bunch of single-node evals at once to deeply understand model performance. Modal Clusters let us scale up to hundreds of B300s with RDMA, then spin them down when we’re done, with minimal overhead. We only pay for what we use, and the team gets to think about how to build intelligence for NEO instead of GPUs and where to find them.”","Sam Sinha, Head of World Model Lab, 1x","Runway ：https://runway.com is building real-world intelligence for the world model era. Their latest frontier video generation and editing models, Gen-4.5 ：https://runway.com/research/introducing-runway-gen-4.5 and Aleph 2.0 ：https://www.youtube.com/watch?v=8vAqhBxCxc0 , generate production-quality video in seconds. Each generation uses Modal Clusters, spreading inference across multiple nodes, and Modal’s autoscaler scales those clusters as a unit across our global compute pool.","“Multi-node inference is a challenging infrastructure problem. It makes elasticity hard, which is table stakes for inference workloads. Modal’s platform, with out of the box features like clustering and configurable autoscaling, has helped us rapidly scale multiple production inference workloads to meet growing demand.”","Modal Clusters are available to all workspaces today. Whether you’re scaling out inference or post-training a frontier model, all it takes is one line of code to scale out your Modal App. Cluster size is bounded by your plan’s GPU limits, so if you’re planning something big, reach out ：https://modal.com/slack .","We can’t wait to see what you build."],"articleImages":[{"sourceUrl":"https://modal-cdn.com/blog/images/decagon-clusters-ga.webp","alt":"Stacked Decagon chat windows for several brands, with an agent asking a customer how it can help","afterParagraph":16,"url":"/media/articles/zbprvsawadktpi8tcqpp6gw4d/6c517f939e5105b8.webp"}],"mediaStatus":"ok","articleBodyZh":["组织需要拥有自己的智能才能取得成功。然而，在大规模训练和服务这种智能时，需要跨多个节点使用每秒千万亿次浮点运算（petaFLOP/s）的计算能力和每秒万亿位（terabit/s）的网络传输。但拥有智能并不意味着必须拥有这些硬件。","在过去的1.5年里，我们一直在实战测试一种新的原语：Modal 集群。今天，我们很高兴地宣布，它们已通过单个装饰器 @modal.clustered 正式开放使用：","Modal 集群可以无缝集成所有现有的 Modal 原语。将你的检查点写入卷：/docs/guide/volumes，通过云存储挂载加载数据：/docs/guide/cloud-bucket-mounts，并使用队列协调你的作业：/docs/guide/queues。在现有集成的基础上，你还能获得：","在其他地方，按需集群访问意味着按小时付费或提前预约。而在 Modal 上，你可以随时运行，只为你使用的部分付费。","集群（容器组）是 Modal 堆栈中的一个全新的工作单元。在支持它们时，我们不得不在堆栈的每一层构建新的系统。","Modal 的传统调度器是贪心的。集群中所有节点会不断轮询新任务，调度器检查任务队列中的可用任务，如果适合就分配。调度器会根据放置、碎片和成本进行启发式处理，但从根本上讲，调度器是基于单节点决定节点是否接取任务。","对于多节点调度，单节点视图是不够的——调度的原子单位从单节点扩展到 N 节点。这意味着我们必须构建一个全新的调度器，具有明确的观察、计划和行动循环：https://en.wikipedia.org/wiki/OODA_loop：观察整个集群，决定每个待处理集群的放置，然后执行。这也是 Modal 其他调度系统的发展方向。在每次迭代中，集群调度器会：","集群调度器从与 Modal 其他部分相同的容量池中获取资源，允许在秒级时间内获取集群。这是市场上获取多节点集群速度最快的方式，也是唯一真正按无服务器价格付费的地方。","多节点工作负载传输大量字节。GLM 4.7 的每个训练步骤都必须将约 717 GB 的 BF16 权重从训练器同步到 rollout 引擎：在 50 Gbps TCP 下几乎需要两分钟，在 RDMA 下不到两秒，且需要重复数千次。对于推理，Llama 3.1 70B 的 prefill-decode 分离每个 32k 令牌的提示需要移动约 10 GB 的 KV 缓存，并且必须在几百毫秒的首令牌时间预算内到达。","RDMA（远程直接内存访问）允许 GPU 或 CPU 内存通过网络在主机之间传输，而无需经过内核空间。跳过软件层使 RDMA 可以达到高达 6.4 Tbps 的速度！这种能力对于大多数多节点工作负载至关重要——训练需要它来进行 DDP all-reduce 和各种模型并行；推理在进行 PD 分离时也需要它来传输 KV 缓存。","RDMA 的支持通常非常困难。每个云提供商的实现略有不同，需要不同的驱动程序、环境变量和用户空间库。","我们的目标非常雄心勃勃：通过一个单一的标志 rdma=True 隐藏所有这些复杂性。如今，你只需设置这个标志，任何使用 PyTorch 和 NCCL 的工作负载就可以开箱即用。我们已经在多个云环境中自动化了 RDMA 设置，使你能够在不产生任何麻烦的情况下深度访问按需集群容量。","这种集成仍然不足以满足我们的需求。由于 Modal 运行的是多租户计算环境，我们更喜欢安全的 gVisor 容器运行时而不是 runc。然而，它不支持 RDMA，因此我们自行将其构建到 gVisor 中：https://github.com/google/gvisor/pull/13964。我们已将这些更改提交到上游，以便开源生态系统的其他人也能访问安全的 RDMA。","我们为基于多节点集群构建的前沿研究和生产应用感到自豪。","Decagon：https://decagon.ai 构建用于客户体验的 AI 代理。Decagon 与 Modal 合作，使用 Miles：https://github.com/radixark/miles 在 Modal Clusters 上对拥有高达一万亿参数的开源模型进行了微调，使其满足客户交互的需求，重点是响应质量和推理效率。作为此次合作的一部分，Modal 为 Miles 提供了 LoRA 支持：https://github.com/radixark/miles/pull/1361，并帮助建立了稳定的开源训练方案。","“模型的性能对于 Decagon 的客户体验至关重要。我们的代理需要理解每个业务的细微差别，并可靠地处理复杂的客户对话。微调使我们能够根据这些需求塑造模型行为，但在万亿参数规模上进行微调是一个巨大的基础设施挑战。Modal Clusters 帮助我们在该规模上进行训练，因此我们可以专注于提高代理的质量和可靠性。”","赛勒斯·阿斯加里，研究主管，Decagon","1x：https://www.1x.tech 正在构建家用机器人 NEO。在 NEO 背后是一个世界模型，1x 在 Modal Clusters 上对其进行预训练：大规模运行使用具有 RDMA 的多节点 B300 集群，评估运行则使用单节点作业群，用于判断模型表现如何。同一个调度器管理两者，因此 1x 可以在需要展开实验时瞬时使用数百个 GPU，并在运行之间缩减规模。","“构建 NEO 需要灵活的工作负载；有些日子我们只进行少量多节点运行，而其他日子我们希望一次启动大量单节点评估，以深入了解模型性能。Modal Clusters 允许我们扩展到数百个带 RDMA 的 B300 集群，然后在完成后缩减，开销最小。我们只需为实际使用的资源付费，团队可以专注于构建 NEO 的智能，而不是 GPU 或它们的位置。”","萨姆·辛哈，世界模型实验室负责人，1x","Runway：https://runway.com 正在为世界模型时代构建现实世界智能。他们最新的前沿视频生成和编辑模型，Gen-4.5：https://runway.com/research/introducing-runway-gen-4.5 和 Aleph 2.0：https://www.youtube.com/watch?v=8vAqhBxCxc0，可在数秒内生成制作级视频。每次生成都使用 Modal 集群，将推理分散到多个节点上，而 Modal 的自动扩展器则可以在我们的全球计算池中按单元扩展这些集群。","“多节点推理是一个具有挑战性的基础设施问题。它使弹性变得困难，而弹性对于推理工作负载来说是基本要求。Modal 平台具有开箱即用的功能，如集群和可配置自动扩展，帮助我们快速扩展多个生产推理工作负载以满足不断增长的需求。”","Modal 集群现已向所有工作区开放。无论您是在扩展推理还是在训练后处理前沿模型，只需一行代码即可扩展您的 Modal 应用。集群规模受您计划的 GPU 限制，因此如果您计划进行大规模操作，请联系：https://modal.com/slack。","我们迫不及待想看看您会创造出什么。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Modal 宣布 Modal Clusters 正式可用，开发者可通过 @modal.clustered 获取多节点集群。该服务支持节点间 InfiniBand verbs 通信，带宽可达 6.4 Tbps，并自动配置 PyTorch 与 NCCL。","background":"Modal 表示其已对这一集群原语测试约 1.5 年。为支持多节点任务，公司新建了按集群整体进行观察、规划和执行的 gang scheduler，并让集群与 Volumes、Cloud Bucket Mounts 和 Queues 等现有原语集成。","viewpoint":"Aioga 判断：此次发布的重点不仅是用单个装饰器调用多节点资源，也包括将集群纳入 Modal 既有的存储、数据加载和任务编排体系。其技术价值仍需结合实际任务表现加以验证。","implications":"可能影响：按需获取多节点集群并按实际使用付费，可能改变部分任务的资源使用方式，但官方关于数秒获取集群、市场最快及真正无服务器定价的表述属于自身主张，不代表所有场景均能达到相同结果。","nextStep":"后续观察：需要关注 Modal Clusters 在不同规模任务中的实际启动时间、通信表现、调度稳定性与费用情况，并核实自动配置 PyTorch 和 NCCL 后的兼容范围，以及与现有 Modal 原语协同运行的效果。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-10-01T11:40:12.224Z","sourceHash":"e786c40e582e3837","review":{"approved":true,"groundedness":95,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["“开发者”属于合理受众概括，来源原文未明确使用该称谓，但不影响事实准确性。","“技术价值仍需结合实际任务表现加以验证”及后续观察内容属于明确标注的审慎判断和建议，不构成事实错误。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","Modal 官方工程博客"],"translations":{"zh-CN":{"title":"Modal Clusters 正式发布，通过 @modal.clustered 提供多节点 GPU 集群","summary":"Modal 宣布 Modal Clusters 正式可用，通过一个装饰器 @modal.clustered 即可获得多节点集群，节点间经 InfiniBand verbs 通信可达 6.4 Tbps，自动配置 PyTorch 和 NCCL。","category":"行业动态","source":"modal.com","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clusters 正式发布，通过 @modal.clustered 提供多节点 GPU 集群 - Aioga AI资讯","description":"Modal 宣布 Modal Clusters 正式可用，通过一个装饰器 @modal.clustered 即可获得多节点集群，节点间经 InfiniBand verbs 通信可达 6.4 Tbps，自动配置 PyTorch 和 NCCL。","url":"https://www.aioga.com/news/zbprvsawadktpi8tcqpp6gw4d/","articleBody":["组织需要拥有自己的智能才能取得成功。然而，在大规模训练和服务这种智能时，需要跨多个节点使用每秒千万亿次浮点运算（petaFLOP/s）的计算能力和每秒万亿位（terabit/s）的网络传输。但拥有智能并不意味着必须拥有这些硬件。","在过去的1.5年里，我们一直在实战测试一种新的原语：Modal 集群。今天，我们很高兴地宣布，它们已通过单个装饰器 @modal.clustered 正式开放使用：","Modal 集群可以无缝集成所有现有的 Modal 原语。将你的检查点写入卷：/docs/guide/volumes，通过云存储挂载加载数据：/docs/guide/cloud-bucket-mounts，并使用队列协调你的作业：/docs/guide/queues。在现有集成的基础上，你还能获得：","在其他地方，按需集群访问意味着按小时付费或提前预约。而在 Modal 上，你可以随时运行，只为你使用的部分付费。","集群（容器组）是 Modal 堆栈中的一个全新的工作单元。在支持它们时，我们不得不在堆栈的每一层构建新的系统。","Modal 的传统调度器是贪心的。集群中所有节点会不断轮询新任务，调度器检查任务队列中的可用任务，如果适合就分配。调度器会根据放置、碎片和成本进行启发式处理，但从根本上讲，调度器是基于单节点决定节点是否接取任务。","对于多节点调度，单节点视图是不够的——调度的原子单位从单节点扩展到 N 节点。这意味着我们必须构建一个全新的调度器，具有明确的观察、计划和行动循环：https://en.wikipedia.org/wiki/OODA_loop：观察整个集群，决定每个待处理集群的放置，然后执行。这也是 Modal 其他调度系统的发展方向。在每次迭代中，集群调度器会：","集群调度器从与 Modal 其他部分相同的容量池中获取资源，允许在秒级时间内获取集群。这是市场上获取多节点集群速度最快的方式，也是唯一真正按无服务器价格付费的地方。","多节点工作负载传输大量字节。GLM 4.7 的每个训练步骤都必须将约 717 GB 的 BF16 权重从训练器同步到 rollout 引擎：在 50 Gbps TCP 下几乎需要两分钟，在 RDMA 下不到两秒，且需要重复数千次。对于推理，Llama 3.1 70B 的 prefill-decode 分离每个 32k 令牌的提示需要移动约 10 GB 的 KV 缓存，并且必须在几百毫秒的首令牌时间预算内到达。","RDMA（远程直接内存访问）允许 GPU 或 CPU 内存通过网络在主机之间传输，而无需经过内核空间。跳过软件层使 RDMA 可以达到高达 6.4 Tbps 的速度！这种能力对于大多数多节点工作负载至关重要——训练需要它来进行 DDP all-reduce 和各种模型并行；推理在进行 PD 分离时也需要它来传输 KV 缓存。","RDMA 的支持通常非常困难。每个云提供商的实现略有不同，需要不同的驱动程序、环境变量和用户空间库。","我们的目标非常雄心勃勃：通过一个单一的标志 rdma=True 隐藏所有这些复杂性。如今，你只需设置这个标志，任何使用 PyTorch 和 NCCL 的工作负载就可以开箱即用。我们已经在多个云环境中自动化了 RDMA 设置，使你能够在不产生任何麻烦的情况下深度访问按需集群容量。","这种集成仍然不足以满足我们的需求。由于 Modal 运行的是多租户计算环境，我们更喜欢安全的 gVisor 容器运行时而不是 runc。然而，它不支持 RDMA，因此我们自行将其构建到 gVisor 中：https://github.com/google/gvisor/pull/13964。我们已将这些更改提交到上游，以便开源生态系统的其他人也能访问安全的 RDMA。","我们为基于多节点集群构建的前沿研究和生产应用感到自豪。","Decagon：https://decagon.ai 构建用于客户体验的 AI 代理。Decagon 与 Modal 合作，使用 Miles：https://github.com/radixark/miles 在 Modal Clusters 上对拥有高达一万亿参数的开源模型进行了微调，使其满足客户交互的需求，重点是响应质量和推理效率。作为此次合作的一部分，Modal 为 Miles 提供了 LoRA 支持：https://github.com/radixark/miles/pull/1361，并帮助建立了稳定的开源训练方案。","“模型的性能对于 Decagon 的客户体验至关重要。我们的代理需要理解每个业务的细微差别，并可靠地处理复杂的客户对话。微调使我们能够根据这些需求塑造模型行为，但在万亿参数规模上进行微调是一个巨大的基础设施挑战。Modal Clusters 帮助我们在该规模上进行训练，因此我们可以专注于提高代理的质量和可靠性。”","赛勒斯·阿斯加里，研究主管，Decagon","1x：https://www.1x.tech 正在构建家用机器人 NEO。在 NEO 背后是一个世界模型，1x 在 Modal Clusters 上对其进行预训练：大规模运行使用具有 RDMA 的多节点 B300 集群，评估运行则使用单节点作业群，用于判断模型表现如何。同一个调度器管理两者，因此 1x 可以在需要展开实验时瞬时使用数百个 GPU，并在运行之间缩减规模。","“构建 NEO 需要灵活的工作负载；有些日子我们只进行少量多节点运行，而其他日子我们希望一次启动大量单节点评估，以深入了解模型性能。Modal Clusters 允许我们扩展到数百个带 RDMA 的 B300 集群，然后在完成后缩减，开销最小。我们只需为实际使用的资源付费，团队可以专注于构建 NEO 的智能，而不是 GPU 或它们的位置。”","萨姆·辛哈，世界模型实验室负责人，1x","Runway：https://runway.com 正在为世界模型时代构建现实世界智能。他们最新的前沿视频生成和编辑模型，Gen-4.5：https://runway.com/research/introducing-runway-gen-4.5 和 Aleph 2.0：https://www.youtube.com/watch?v=8vAqhBxCxc0，可在数秒内生成制作级视频。每次生成都使用 Modal 集群，将推理分散到多个节点上，而 Modal 的自动扩展器则可以在我们的全球计算池中按单元扩展这些集群。","“多节点推理是一个具有挑战性的基础设施问题。它使弹性变得困难，而弹性对于推理工作负载来说是基本要求。Modal 平台具有开箱即用的功能，如集群和可配置自动扩展，帮助我们快速扩展多个生产推理工作负载以满足不断增长的需求。”","Modal 集群现已向所有工作区开放。无论您是在扩展推理还是在训练后处理前沿模型，只需一行代码即可扩展您的 Modal 应用。集群规模受您计划的 GPU 限制，因此如果您计划进行大规模操作，请联系：https://modal.com/slack。","我们迫不及待想看看您会创造出什么。"]},"en":{"title":"Modal Clusters have been officially launched, providing multi-node GPU clusters through @modal.clustered","summary":"Modal announced that Modal Clusters are officially available. With a decorator @modal.clustered, multi-node clusters can be obtained, with up to 6.4 Tbps of communication between nodes via InfiniBand verbs, and automatic configuration of PyTorch and NCCL.","category":"Industry","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clusters have been officially launched, providing multi-node GPU clusters through @modal.clustered - Aioga AI News","description":"Modal announced that Modal Clusters are officially available. With a decorator @modal.clustered, multi-node clusters can be obtained, with up to 6.4 Tbps of communication between n...","url":"https://www.aioga.com/en/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:02:51.981Z"},"ja":{"title":"Modal Clastesは正式にローンチし、@modal.clusteredを通じてマルチノードGPUクラスターを提供しています","summary":"ModarはModar Clastesが正式に利用可能であると発表しました。@modal.clusteredというデコレーターを使えば、多ノードクラスターを実現でき、Tஇஇன்фиni Band動詞による最大6.4 Tbpsのノード通信や、PyTorchおよびNCCLの自動設定が可能です。","category":"業界動向","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastesは正式にローンチし、@modal.clusteredを通じてマルチノードGPUクラスターを提供しています - Aioga AIニュース","description":"ModarはModar Clastesが正式に利用可能であると発表しました。@modal.clusteredというデコレーターを使えば、多ノードクラスターを実現でき、Tஇஇன்фиni Band動詞による最大6.4 Tbpsのノード通信や、PyTorchおよびNCCLの自動設定が可能です。","url":"https://www.aioga.com/ja/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:02:52.996Z"},"ko":{"title":"Modal Clastes가 공식적으로 출시되었으며, @modal.clustered를 통해 다중 노드 GPU 클러스터를 제공합니다","summary":"Modar는 Modar Clastes가 공식적으로 출시되었다고 발표했습니다. @modal.clustered라는 데코레이터를 사용하면 다중 노드 클러스터를 구현할 수 있으며, Tஇஇன்фиni Band 동사를 통한 최대 6.4 Tbps의 노드 통신과 PyTorch와 NCCL의 자동 구성 등이 가능합니다.","category":"업계 동향","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes가 공식적으로 출시되었으며, @modal.clustered를 통해 다중 노드 GPU 클러스터를 제공합니다 - Aioga AI 뉴스","description":"Modar는 Modar Clastes가 공식적으로 출시되었다고 발표했습니다. @modal.clustered라는 데코레이터를 사용하면 다중 노드 클러스터를 구현할 수 있으며, Tஇஇன்фиni Band 동사를 통한 최대 6.4 Tbps의 노드 통신과 PyTorch와 NCCL의 자동 구성 등이 가능합니다.","url":"https://www.aioga.com/ko/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:02:52.947Z"},"es":{"title":"Modal Clastes se ha lanzado oficialmente, ofreciendo clústeres de GPU multinodo a través de @modal.clustered","summary":"Modar anunció que Modar Clastes está oficialmente disponible. Con un decorador llamado @modal.clustered, se pueden lograr clústeres multinodo, con hasta 6,4 Tbps de comunicación nodal mediante verbos en banda Tஇஇஇன்фиni, y configuración automática de PyTorch y NCCL.","category":"Industria","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes se ha lanzado oficialmente, ofreciendo clústeres de GPU multinodo a través de @modal.clustered - Aioga Noticias de IA","description":"Modar anunció que Modar Clastes está oficialmente disponible. Con un decorador llamado @modal.clustered, se pueden lograr clústeres multinodo, con hasta 6,4 Tbps de comunicación no...","url":"https://www.aioga.com/es/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:16.872Z"},"fr":{"title":"Modal Clastes a officiellement été lancé, proposant des clusters GPU multi-nœuds via @modal.clustered","summary":"Modar a annoncé que Modar Clastes est officiellement disponible. Avec un décorateur appelé @modal.clustered, des clusters multi-nœuds peuvent être réalisés, avec jusqu’à 6,4 Tbps de communication par nœuds via Tஇஇன்фиni en bandes, et une configuration automatique de PyTorch et NCCL.","category":"Industrie","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes a officiellement été lancé, proposant des clusters GPU multi-nœuds via @modal.clustered - Aioga Actualités IA","description":"Modar a annoncé que Modar Clastes est officiellement disponible. Avec un décorateur appelé @modal.clustered, des clusters multi-nœuds peuvent être réalisés, avec jusqu’à 6,4 Tbps d...","url":"https://www.aioga.com/fr/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:17.385Z"},"de":{"title":"Modal Clastes ist offiziell gestartet und bietet Multi-Node-GPU-Cluster über @modal.clustered an","summary":"Modar gab bekannt, dass Modar Clastes offiziell verfügbar ist. Mit einem Decorator namens @modal.clustered können Mehrknoten-Cluster erreicht werden, mit bis zu 6,4 Tbps Knotenkommunikation über Tஇஇன்фиni Bandverbs sowie automatischer Konfiguration von PyTorch und NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes ist offiziell gestartet und bietet Multi-Node-GPU-Cluster über @modal.clustered an - Aioga KI-News","description":"Modar gab bekannt, dass Modar Clastes offiziell verfügbar ist. Mit einem Decorator namens @modal.clustered können Mehrknoten-Cluster erreicht werden, mit bis zu 6,4 Tbps Knotenkomm...","url":"https://www.aioga.com/de/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:17.238Z"},"pt-BR":{"title":"Modal Clastes foi oficialmente lançado, oferecendo clusters de GPU multi-nós através do @modal.clustered","summary":"A Modar anunciou que o Modar Clastes está oficialmente disponível. Com um decorador chamado @modal.clustered, é possível realizar clusters multi-nós, com até 6,4 Tbps de comunicação de nós via verbos na banda Tஇஇஇன்фиni, e configuração automática do PyTorch e NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes foi oficialmente lançado, oferecendo clusters de GPU multi-nós através do @modal.clustered - Aioga Notícias de IA","description":"A Modar anunciou que o Modar Clastes está oficialmente disponível. Com um decorador chamado @modal.clustered, é possível realizar clusters multi-nós, com até 6,4 Tbps de comunicaçã...","url":"https://www.aioga.com/pt-BR/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:41.900Z"},"ru":{"title":"Modal Clastes официально запустилась, предлагая многоузловые кластеры GPU через @modal.clustered","summary":"Modar объявила, что Modar Clastes официально доступен. С помощью декоратора под названием @modal.clustered можно создать многоузловые кластеры с частотой до 6,4 Тбит/с связи узлов через глаголы Tஇன்фиni Band и автоматической конфигурацией PyTorch и NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes официально запустилась, предлагая многоузловые кластеры GPU через @modal.clustered - Aioga Новости ИИ","description":"Modar объявила, что Modar Clastes официально доступен. С помощью декоратора под названием @modal.clustered можно создать многоузловые кластеры с частотой до 6,4 Тбит/с связи узлов...","url":"https://www.aioga.com/ru/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:41.568Z"},"ar":{"title":"تم إطلاق مودال كلاستس رسميا، مقدما عناقيد بطاقات رسومات متعددة العقد عبر @modal.clustered","summary":"أعلنت مودار أن مودار كلاستس أصبحت متاحة رسميا. مع ديكور يسمى @modal.clustered، يمكن تحقيق مجموعات متعددة العقد، مع اتصال يصل إلى 6.4 تيرابايت في الثانية عبر أفعال Tஇஇன்фиni Band من خلال أفعال Tஇஇன்фиni، وتكوين تلقائي ل PyTorch وNCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"تم إطلاق مودال كلاستس رسميا، مقدما عناقيد بطاقات رسومات متعددة العقد عبر @modal.clustered - Aioga أخبار الذكاء الاصطناعي","description":"أعلنت مودار أن مودار كلاستس أصبحت متاحة رسميا. مع ديكور يسمى @modal.clustered، يمكن تحقيق مجموعات متعددة العقد، مع اتصال يصل إلى 6.4 تيرابايت في الثانية عبر أفعال Tஇஇன்фиni Band من...","url":"https://www.aioga.com/ar/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:41.843Z"},"hi":{"title":"मोडल क्लैस्टेस आधिकारिक तौर पर लॉन्च हो गया है, जो @modal.clustered के माध्यम से मल्टी-नोड GPU क्लस्टर पेश करता है","summary":"मोदार ने घोषणा की कि मोदार क्लैस्टेस आधिकारिक तौर पर उपलब्ध है। @modal.क्लस्टर नामक डेकोरेटर के साथ, मल्टी-नोड क्लस्टर प्राप्त किए जा सकते हैं, जिसमें Tஇஇன்фиni बैंड क्रियाओं के माध्यम से 6.4 Tbps तक नोड संचार और PyTorch और NCCL का स्वचालित कॉन्फ़िगरेशन होता है।","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"मोडल क्लैस्टेस आधिकारिक तौर पर लॉन्च हो गया है, जो @modal.clustered के माध्यम से मल्टी-नोड GPU क्लस्टर पेश करता है - Aioga AI समाचार","description":"मोदार ने घोषणा की कि मोदार क्लैस्टेस आधिकारिक तौर पर उपलब्ध है। @modal.क्लस्टर नामक डेकोरेटर के साथ, मल्टी-नोड क्लस्टर प्राप्त किए जा सकते हैं, जिसमें Tஇஇன்фиni बैंड क्रियाओं के मा...","url":"https://www.aioga.com/hi/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:05.591Z"},"it":{"title":"Modal Clastes è stato ufficialmente lanciato, offrendo cluster GPU multi-nodo tramite @modal.clustered","summary":"Modar ha annunciato che Modar Clastes è ufficialmente disponibile. Con un decorator chiamato @modal.clustered, si possono realizzare cluster multi-nodo, con fino a 6,4 Tbps di comunicazione tra nodi tramite verbi Tஇஇஇன்фиni Band e configurazione automatica di PyTorch e NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes è stato ufficialmente lanciato, offrendo cluster GPU multi-nodo tramite @modal.clustered - Aioga Notizie IA","description":"Modar ha annunciato che Modar Clastes è ufficialmente disponibile. Con un decorator chiamato @modal.clustered, si possono realizzare cluster multi-nodo, con fino a 6,4 Tbps di comu...","url":"https://www.aioga.com/it/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:55.387Z"},"nl":{"title":"Modal Clusters officieel gelanceerd, biedt multi-node GPU-clusters via @modal.clustered","summary":"Modal heeft aangekondigd dat Modal Clusters nu algemeen beschikbaar is. Met één decorator, @modal.clustered, krijgt men toegang tot multi-node clusters, waarbij communicatie tussen nodes via InfiniBand verbs snelheden tot 6.4 Tbps bereikt en PyTorch en NCCL automatisch worden geconfigureerd.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clusters officieel gelanceerd, biedt multi-node GPU-clusters via @modal.clustered - Aioga AI-nieuws","description":"Modal heeft aangekondigd dat Modal Clusters nu algemeen beschikbaar is. Met één decorator, @modal.clustered, krijgt men toegang tot multi-node clusters, waarbij communicatie tussen...","url":"https://www.aioga.com/nl/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:03:56.117Z"},"tr":{"title":"Modal Clastes resmi olarak piyasaya sürüldü ve @modal.clustered üzerinden çok düğümlü GPU kümeleri sunuyor","summary":"Modar, Modar Clastes'in resmi olarak kullanıma sunulduğunu duyurdu. @modal.clustered adlı bir dekoratör ile, Tஇஇன்фиni Bant fiilleri üzerinden 6,4 Tbps'ye kadar düğüm iletişimi ve PyTorch ile NCCL'nin otomatik yapılandırması ile çoklu düğümlü kümeler elde edilebilir.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes resmi olarak piyasaya sürüldü ve @modal.clustered üzerinden çok düğümlü GPU kümeleri sunuyor - Aioga AI Haberleri","description":"Modar, Modar Clastes'in resmi olarak kullanıma sunulduğunu duyurdu. @modal.clustered adlı bir dekoratör ile, Tஇஇன்фиni Bant fiilleri üzerinden 6,4 Tbps'ye kadar düğüm iletişimi ve...","url":"https://www.aioga.com/tr/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:30.047Z"},"vi":{"title":"Modal Clastes chính thức ra mắt, cung cấp các cụm GPU đa nút thông qua @modal.clustered","summary":"Modar đã thông báo rằng Modar Clastes đã chính thức có mặt. Với một decorator gọi là @modal.clustered, có thể đạt được các cụm nhiều nút, với tốc độ giao tiếp nút lên đến 6,4 Tbps qua các động từ Band Tஇஇன்фиni, và cấu hình tự động PyTorch và NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes chính thức ra mắt, cung cấp các cụm GPU đa nút thông qua @modal.clustered - Tin tức AI Aioga","description":"Modar đã thông báo rằng Modar Clastes đã chính thức có mặt. Với một decorator gọi là @modal.clustered, có thể đạt được các cụm nhiều nút, với tốc độ giao tiếp nút lên đến 6,4 Tbps...","url":"https://www.aioga.com/vi/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:29.934Z"},"id":{"title":"Modal Clusters resmi diluncurkan, menyediakan klaster GPU multinode melalui @modal.clustered","summary":"Modal mengumumkan bahwa Modal Clusters kini tersedia secara resmi. Klaster multinode dapat diperoleh hanya dengan satu dekorator, @modal.clustered, dengan komunikasi antarnode melalui InfiniBand verbs yang dapat mencapai 6.4 Tbps serta konfigurasi PyTorch dan NCCL secara otomatis.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clusters resmi diluncurkan, menyediakan klaster GPU multinode melalui @modal.clustered - Berita AI Aioga","description":"Modal mengumumkan bahwa Modal Clusters kini tersedia secara resmi. Klaster multinode dapat diperoleh hanya dengan satu dekorator, @modal.clustered, dengan komunikasi antarnode mela...","url":"https://www.aioga.com/id/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:19.347Z"},"th":{"title":"Modal Clastes ได้เปิดตัวอย่างเป็นทางการแล้ว โดยนําเสนอคลัสเตอร์ GPU หลายโหนดผ่าน @modal.clustered","summary":"Modar ประกาศว่า Modar Clastes พร้อมใช้งานอย่างเป็นทางการแล้วด้วยตัวตกแต่งที่ชื่อ @modal.clustered สามารถสร้างคลัสเตอร์หลายโหนดได้ โดยมีการสื่อสารโหนดผ่าน Tஇஇன்фиni Band verbs สูงสุด 6.4 Tbps และการตั้งค่าอัตโนมัติของ PyTorch และ NCCL","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes ได้เปิดตัวอย่างเป็นทางการแล้ว โดยนําเสนอคลัสเตอร์ GPU หลายโหนดผ่าน @modal.clustered - ข่าว AI Aioga","description":"Modar ประกาศว่า Modar Clastes พร้อมใช้งานอย่างเป็นทางการแล้วด้วยตัวตกแต่งที่ชื่อ @modal.clustered สามารถสร้างคลัสเตอร์หลายโหนดได้ โดยมีการสื่อสารโหนดผ่าน Tஇஇன்фиni Band verbs สูงสุ...","url":"https://www.aioga.com/th/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:53.713Z"},"pl":{"title":"Modal Clastes oficjalnie uruchomił swoje usługi, oferując wielowęzłowe klastry GPU poprzez @modal.clustered","summary":"Modar ogłosił, że Modar Clastes jest oficjalnie dostępny. Dzięki dekoratorowi o nazwie @modal.clustered można uzyskać klastry wielowęzłowe, z komunikacją węzłów do 6,4 Tbps za pomocą czasowników Tஇஇன்фиni oraz automatyczną konfiguracją PyTorch i NCCL.","category":"行业动态","source":"Modal 官方工程博客（RSS）","aggregationSource":"Modal 官方工程博客（RSS）","pageTitle":"Modal Clastes oficjalnie uruchomił swoje usługi, oferując wielowęzłowe klastry GPU poprzez @modal.clustered - Aioga Wiadomości AI","description":"Modar ogłosił, że Modar Clastes jest oficjalnie dostępny. Dzięki dekoratorowi o nazwie @modal.clustered można uzyskać klastry wielowęzłowe, z komunikacją węzłów do 6,4 Tbps za pomo...","url":"https://www.aioga.com/pl/news/zbprvsawadktpi8tcqpp6gw4d/","contentTranslated":true,"translationStatus":"translated","translationRetryAt":"","translationError":"","sourceHash":"3d30c3b395f77739","translatedAt":"2026-10-01T08:04:54.008Z"}},"evidenceTier":"verified-news","reviewStatus":"editorial-selected","indexable":true,"editorialCover":""}}