{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-23T08:01:28.298Z","headline":"Google 推出 Tunix：基于 JAX 的高吞吐智能体后训练库","description":"Google 发布 Tunix，一个基于 JAX 的原生后训练库，旨在消除多轮、使用工具的 LLM 推理智能体训练中的 TPU 闲置瓶颈。Tunix 通过高并发异步 rollout 与解耦的生产者-消费者流水线最大化硬件吞吐量，确保训练器持续获得数据。该库提供即插即用抽象和持续宏观级性能分析，便于集成自定义环境。","url":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/","mainEntityOfPage":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/","datePublished":"2026-07-21T16:00:05.337Z","dateModified":"2026-07-21T16:00:05.337Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://developers.googleblog.com/scaling-agentic-rl-high-throughput-agentic-training-with-tunix","https://aihot.virxact.com/items/cmruuc0d80006bii0tzwael16"],"canonicalUrl":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/","directAnswer":{"@type":"Answer","text":"Google 发布 Tunix 最新版本，这是一套基于 JAX 的后训练库，面向多轮决策、工具调用和复杂环境交互的智能体强化学习训练，重点缓解环境步骤导致 TPU 等待和利用率下降的问题。","url":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/","dateCreated":"2026-07-21T16:00:05.337Z","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":"developers.googleblog.com source article","url":"https://developers.googleblog.com/scaling-agentic-rl-high-throughput-agentic-training-with-tunix","datePublished":"2026-07-21T16:00:05.337Z","provider":{"@type":"Organization","name":"developers.googleblog.com","url":"https://developers.googleblog.com/scaling-agentic-rl-high-throughput-agentic-training-with-tunix"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cmruuc0d80006bii0tzwael16","datePublished":"2026-07-21T16:00:05.337Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cmruuc0d80006bii0tzwael16"}}],"aggregationSource":"Google Developers Blog（RSS）","originalPublisher":{"name":"developers.googleblog.com","url":"https://developers.googleblog.com/scaling-agentic-rl-high-throughput-agentic-training-with-tunix"},"article":{"id":"cmruuc0d80006bii0tzwael16","slug":"cmruuc0d80006bii0tzwael16","url":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/","title":"Google 推出 Tunix：基于 JAX 的高吞吐智能体后训练库","title_en":"Scaling Agentic RL： High-Throughput Agentic Training with Tunix","summary":"Google 发布 Tunix，一个基于 JAX 的原生后训练库，旨在消除多轮、使用工具的 LLM 推理智能体训练中的 TPU 闲置瓶颈。Tunix 通过高并发异步 rollout 与解耦的生产者-消费者流水线最大化硬件吞吐量，确保训练器持续获得数据。该库提供即插即用抽象和持续宏观级性能分析，便于集成自定义环境。","source":"Google Developers Blog（RSS）","sourceUrl":"https://developers.googleblog.com/scaling-agentic-rl-high-throughput-agentic-training-with-tunix","aiHotUrl":"https://aihot.virxact.com/items/cmruuc0d80006bii0tzwael16","publishedAt":"2026-07-21T16:00:05.337Z","category":"产品更新","score":63,"selected":true,"articleBody":["The focus of LLM alignment has rapidly shifted from static chatbot alignment to dynamic agentic workflows . Today’s models don't just talk—they execute multi-step reasoning, call external APIs, and interact with complex environments.","Training reasoning agents encounters special challenges and bottlenecks. The recent evolution of agentic RL training shifts the process from single-turn alignment to multi-turn decision-making with complex environment interactions and tool usage. This shift raises new challenges on the infrastructure side for rollout performance and efficiency; when an agent pauses to execute code, query a database, or wait on a web search, the expensive AI accelerator utilization plummets as TPUs sit idle waiting for environment steps.","Tunix —Google’s post-training library—natively solves this bottleneck in its latest release, introducing an efficient, composable framework for training LLM agents at scale. Tunix keeps accelerators fully utilized on two fronts:","Beyond orchestration, agentic RL requires specialized observability. While standard profilers like XProf offer deep operator-level traces, their high overhead limits them to short, sporadic captures. Tunix introduces continuous, lightweight instrumentation built directly around domain-specific RL metrics. By correlating these high-level loop metrics with TPU timelines, developers get a global view of execution efficiency to quickly spot and resolve system bottlenecks.","Ultimately, Tunix is built to maximize TPU throughput, keep environments modular, and make multi-turn training efficiency fully transparent. Here is how it works under the hood.","Achieving peak hardware throughput means keeping TPUs constantly busy. Tunix accomplishes this by combining asynchronous rollouts to eliminate execution bubbles and stragglers with a decoupled pipeline that continuously streams data to the trainer.","In Agentic RL, trajectory generation (rollout) is the most time-consuming phase. However, the traditional synchronous rollout architecture creates two major problems, as depicted in the figure below.","Tunix solves this with an Asynchronous Trajectory Collector Engine .","This architecture completely overlaps model inference, tool execution, and reward computation, preserving high hardware utilization.","While async rollouts solve trajectory generation bottlenecks, another critical challenge for hardware efficiency in the end-to-end RL workflow is bridging dynamic, variable-length, and potentially long-tail rollouts with a strictly synchronous training loop. A naive approach relies on a synchronization point that forces the accelerator to wait until an entire batch of trajectories is complete before initiating the training step, starving the trainer TPU. Tunix eliminates this bottleneck by decoupling rollout and training into a continuous producer-consumer pipeline (illustrated in the diagram below):","The moment a trajectory group is complete, it is post-processed, scored, and streamed directly into the trainer. This pipeline ensures the synchronous trainer is constantly fed, maximizing end-to-end throughput.","A major friction point in RL frameworks is the rigid coupling of the algorithm to the environment loop. Modifying a codebase to support a new open-source software (OSS) benchmark like SWE-bench, WebArena, or a custom game engine often requires a massive rewrite.","Tunix resolves this with a decoupled, composable architecture. By exposing a clean API boundary, Tunix automates step invocation and lifecycle management so you can focus entirely on core interaction logic.","Why it matters: You can onboard any open-source RL environment in minutes. Because the agent and environment logic are completely decoupled from the training workflow, swapping a single-turn math verifier for an interactive bash terminal requires zero modifications to your training code. To demonstrate the power of this composable design, we next showcase a few examples of how easily new agents, models, or environments can be used. You can find more detailed examples of customized Agent/Env in our recipes：https://github.com/google/tunix/tree/main/examples.","Tunix offers built-in classes like ModelAgent and ToolAgent that work immediately via configuration.","Alternatively, you can build your own custom Agent and add specific logic on how to process the model responses. Tunix will automatically wire this agent in the end to end training workflow. E.g. SWEAgent , FrozenLakeAgent","Similar to Agents, Tunix offers a number of pre-built environments including TaskEnvironment , ToolEnvironment . Alternatively, you can also bring your own custom environment by simply implementing a few main APIs, including any open source environment such as the Gymnasium example below.","When running asynchronous agentic training at scale, traditional logging falls short. You need granular yet domain-specific visibility to identify efficiency problems: Is the bottleneck in the generation phase? Is the tool call taking too much time? Or is the data-loader too slow?","Standard profilers like XProf：https://openxla.org/xprof provide detailed, op-level traces to understand micro-level performance like kernel and model execution. However, capturing long-spanning traces with these tools is typically cost-prohibitive, and identifying macro-level bottlenecks within the noise of low-level data remains difficult. For the complex workflows of agentic RL, developers need a lightweight, macro-level view built on domain-specific metrics that map directly to RL stages.","Tunix delivers this big picture by carefully tracking a minimal set of critical RL-specific metrics that represent both the global pipeline (how rollout, training, and weight sync phases interact) and important sub-steps (each model call, environment interaction, etc.). Because they are lightweight, these metrics run continuously throughout the entire training job. Users can quickly identify where the workflow is stalling globally, and then deploy a tool like XProf for targeted, further debugging.","The figure above illustrates a Perfetto trace captured from a multi-turn agentic training job, detailing the staged execution timelines across CPU threads and TPU devices. As demonstrated by the trace, TPU device utilization is much higher than that of the CPU threads, whose idle time is primarily due to environment execution latency.","This macro-level tracing of staged RL pipelines enables you to:","Tunix turns the \"black box\" of distributed multi-turn RL into a transparent, optimizable timeline for the training job.","If you are evaluating frameworks for Agentic RL, here is how Tunix stands out:","Whether you are reproducing SOTA reasoning models, fine-tuning the Gemma or Qwen families to \"think,\" or deploying complex multi-agent systems, Tunix provides the high-performance foundation needed for the next generation of reasoning agents. Start building today!","Tunix is under active open-source development by Google and the wider community. If you are building the next generation of reasoning agents, drop by our GitHub Issues and let us know what environments you are plugging in."],"articleImages":[{"sourceUrl":"https://storage.googleapis.com/gweb-developer-goog-blog-cms-assets/site/20260519-162827/images/g-dev.svg","alt":"Google for Developers","afterParagraph":0,"url":"/media/articles/cmruuc0d80006bii0tzwael16/8d66d7423366f696.jpg"},{"sourceUrl":"https://storage.googleapis.com/gweb-developer-goog-blog-assets/images/Screenshot_2026-07-13_at_10.15.53AM.original.png","alt":"Screenshot 2026-07-13 at 10.15.53 AM","afterParagraph":6,"url":"/media/articles/cmruuc0d80006bii0tzwael16/c8684f46db05bf19.png"},{"sourceUrl":"https://storage.googleapis.com/gweb-developer-goog-blog-assets/images/Screenshot_2026-07-13_at_10.16.08AM.original.png","alt":"Screenshot 2026-07-13 at 10.16.08 AM","afterParagraph":9,"url":"/media/articles/cmruuc0d80006bii0tzwael16/13958a2983cc54fb.png"},{"sourceUrl":"https://storage.googleapis.com/gweb-developer-goog-blog-assets/images/blog-perfetto.original.png","alt":"blog-perfetto","afterParagraph":19,"url":"/media/articles/cmruuc0d80006bii0tzwael16/f82ddc254179b750.png"}],"mediaStatus":"ok","articleBodyZh":["大型语言模型（LLM）对齐的重点已迅速从静态聊天机器人对齐转向动态自治工作流。如今的模型不仅会对话——它们还能执行多步推理、调用外部 API 并与复杂环境互动。","训练推理型智能体面临特殊的挑战和瓶颈。智能体强化学习（agentic RL）训练的最新发展将过程从单轮对齐转向多轮决策，其中涉及复杂的环境互动和工具使用。这一转变在基础设施方面提出了新的挑战，影响展开性能和效率；当智能体暂停执行代码、查询数据库或等待网络搜索时，昂贵的 AI 加速器利用率会急剧下降，因为 TPU 在等待环境步骤时处于空闲状态。","Tunix——谷歌的后训练库——在其最新发布版本中原生地解决了这一瓶颈，引入了一个高效、可组合的框架，用于大规模训练 LLM 智能体。Tunix 在两个方面保持加速器的充分利用：","除了编排之外，智能体强化学习还需要专门的可观测性。虽然标准分析器如 XProf 提供深度操作级别的轨迹，但其高开销限制了只能进行短时、零散的捕获。Tunix 引入了围绕特定领域 RL 指标直接构建的连续轻量级检测工具。通过将这些高层循环指标与 TPU 时间线关联，开发者可以获得执行效率的全局视图，从而快速发现并解决系统瓶颈。","最终，Tunix 的设计目标是最大化 TPU 吞吐量，保持环境模块化，并使多轮训练效率完全透明。其内部工作机制如下。","实现硬件峰值吞吐量意味着需要让 TPU 不断保持繁忙状态。Tunix 通过结合异步 rollout 以消除执行空洞和滞后者，以及一个连续向训练器流式传输数据的解耦流水线来实现这一目标。","在智能体强化学习中，轨迹生成（rollout）是最耗时的阶段。然而，传统的同步 rollout 架构会产生两个主要问题，如下图所示。","Tunix 通过异步轨迹收集引擎（Asynchronous Trajectory Collector Engine）解决了这个问题。","这种架构将模型推理、工具执行和奖励计算完全重叠，从而保持高硬件利用率。","尽管异步 rollout 解决了轨迹生成的瓶颈，但端到端 RL 工作流中硬件效率的另一个关键挑战是将动态、长度可变并且可能是长尾的 rollout 与严格同步的训练循环相连接。一个简单的做法是依赖一个同步点，这会强制加速器等待整个批次的轨迹完成后才能启动训练步骤，从而使训练 TPU 处于空闲状态。Tunix 通过将 rollout 和训练解耦成连续的生产者-消费者管道（如下图所示）消除了这一瓶颈：","轨迹组一旦完成，就会被后处理、评分，并直接流入训练器。这个管道确保同步训练器不断得到数据输入，从而最大化端到端吞吐量。","RL 框架中的一个主要摩擦点是算法与环境循环的严格耦合。修改代码库以支持新的开源软件（OSS）基准，例如 SWE-bench、WebArena 或自定义游戏引擎，通常需要大规模重写。","Tunix 通过解耦、可组合的架构解决了这个问题。通过公开干净的 API 边界，Tunix 自动化步骤调用和生命周期管理，使你可以完全专注于核心交互逻辑。","重要性：你可以在几分钟内接入任何开源 RL 环境。由于代理和环境逻辑与训练工作流完全解耦，用一个单步数学验证器替换为交互式 bash 终端无需修改任何训练代码。为了展示这种可组合设计的强大功能，我们接下来展示几个新代理、模型或环境能够多么轻松地使用的示例。你可以在我们的示例配方中找到更多定制化 Agent/Env 的详细示例：https://github.com/google/tunix/tree/main/examples。","Tunix 提供了内置类，如 ModelAgent 和 ToolAgent，可通过配置立即工作。","或者，您可以构建自己的自定义Agent，并添加具体逻辑来处理模型响应。Tunix会在端到端训练工作流中自动连接此Agent。例如：SWEAgent，FrozenLakeAgent","类似于Agents，Tunix提供了许多预构建的环境，包括TaskEnvironment，ToolEnvironment。或者，您也可以通过简单地实现几个主要API来引入自己的自定义环境，包括任何开源环境，例如下面的Gymnasium示例。","在大规模运行异步agentic训练时，传统日志记录显得不足。您需要细粒度但特定领域的可见性来识别效率问题：瓶颈出现在生成阶段吗？工具调用耗时过多吗？还是数据加载器过慢？","像XProf这样的标准分析器：https://openxla.org/xprof 提供详细的操作级追踪，以了解如内核和模型执行等微观性能。然而，使用这些工具捕获长跨度追踪通常成本高昂，并且在低级数据噪声中识别宏观瓶颈仍然困难。对于agentic RL的复杂工作流，开发者需要一种基于领域特定指标的轻量级宏观视图，这些指标能够直接映射到RL阶段。","Tunix通过仔细跟踪一组关键的RL特定指标提供了这一整体视图，这些指标既代表全球管道（如rollout、训练和权重同步阶段的交互），也代表重要子步骤（每次模型调用、环境交互等）。由于这些指标轻量，它们可以在整个训练任务中持续运行。用户可以快速识别工作流的全局停滞位置，然后部署像XProf这样的工具进行有针对性的进一步调试。","上图显示了从一个多轮agentic训练任务中捕获的Perfetto追踪，详细展示了CPU线程和TPU设备的分阶段执行时间线。如追踪所示，TPU设备的利用率远高于CPU线程，其空闲时间主要由于环境执行延迟。","这种分阶段RL管道的宏观追踪使您能够：","Tunix 将分布式多轮强化学习（RL）的“黑箱”变成了训练任务可透明、可优化的时间线。","如果你正在评估用于智能体强化学习（Agentic RL）的框架，Tunix 的优势如下：","无论你是在复现最先进的推理模型、微调 Gemma 或 Qwen 系列模型以“思考”，还是部署复杂的多智能体系统，Tunix 都提供了下一代推理智能体所需的高性能基础。从今天开始构建吧！","Tunix 正在由 Google 和更广泛的社区积极进行开源开发。如果你正在构建下一代推理智能体，欢迎访问我们的 GitHub Issues，让我们知道你正在接入哪些环境。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Google 发布 Tunix 最新版本，这是一套基于 JAX 的后训练库，面向多轮决策、工具调用和复杂环境交互的智能体强化学习训练，重点缓解环境步骤导致 TPU 等待和利用率下降的问题。","background":"智能体训练已从单轮对齐扩展到多轮决策。智能体执行代码、查询数据库或等待网络搜索时，传统同步 rollout 容易产生执行空档与慢任务拖延，轨迹生成也成为训练流程中最耗时的阶段。","viewpoint":"Aioga 判断，Tunix 的核心价值在于把智能体强化学习的系统瓶颈作为整体工程问题处理：异步轨迹收集负责减少等待，解耦流水线持续向训练器供数，轻量监测则帮助定位效率问题。","implications":"值得关注的是，该方案可能提升多轮智能体训练中的 TPU 吞吐，并让自定义环境保持模块化。不过公开材料未提供具体性能数字，因此目前不能据此判断其相对其他训练框架的提升幅度。","nextStep":"后续值得关注 Tunix 的实际集成方式、异步轨迹收集在不同环境等待模式下的表现，以及持续监测指标如何与 TPU 时间线关联。Aioga 判断，公开基准和复现实验将是评估效果的重要依据。","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":"2026-07-23T03:10:05.079Z","sourceHash":"b24fd73318a2fbc9","review":{"approved":true,"groundedness":96,"clarity":92,"duplicationRisk":18,"blockingIssues":[],"notes":["“Aioga 判断”明确标示为观点，没有将评价性结论冒充来源事实。","“可能提升”及“不能据此判断提升幅度”等表述保留了必要的不确定性，并准确反映公开材料缺少具体性能数字这一限制。","nextStep 中关于公开基准和复现实验重要性的判断属于合理的后续观察建议，并非来源材料中的既有事实。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","low-source-overlap","no-html","independent-ai-review"]}},"tags":["产品更新","Google Developers Blog（RSS）"],"translations":{"zh-CN":{"title":"Google 推出 Tunix：基于 JAX 的高吞吐智能体后训练库","summary":"Google 发布 Tunix，一个基于 JAX 的原生后训练库，旨在消除多轮、使用工具的 LLM 推理智能体训练中的 TPU 闲置瓶颈。Tunix 通过高并发异步 rollout 与解耦的生产者-消费者流水线最大化硬件吞吐量，确保训练器持续获得数据。该库提供即插即用抽象和持续宏观级性能分析，便于集成自定义环境。","category":"产品更新","source":"developers.googleblog.com","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google 推出 Tunix：基于 JAX 的高吞吐智能体后训练库 - Aioga AI资讯","description":"Google 发布 Tunix，一个基于 JAX 的原生后训练库，旨在消除多轮、使用工具的 LLM 推理智能体训练中的 TPU 闲置瓶颈。Tunix 通过高并发异步 rollout 与解耦的生产者-消费者流水线最大化硬件吞吐量，确保训练器持续获得数据。该库提供即插即用抽象和持续宏观级性能分析，便于集成自定义环境。","url":"https://www.aioga.com/news/cmruuc0d80006bii0tzwael16/"},"en":{"title":"Google launches Tunix: a high-throughput agent fine-tuning library based on JAX","summary":"Google released Tunix, a native post-training library based on JAX, designed to eliminate TPU idling bottlenecks in the training of multi-turn, tool-using LLM reasoning agents. Tunix maximizes hardware throughput through high-concurrency asynchronous rollouts and a decoupled producer-consumer pipeline, ensuring that the trainer continuously receives data. The library provides plug-and-play abstractions and continuous macro-level performance analysis, making it easy to integrate custom environments.","category":"Products","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google launches Tunix: a high-throughput agent fine-tuning library based on JAX - Aioga AI News","description":"Google released Tunix, a native post-training library based on JAX, designed to eliminate TPU idling bottlenecks in the training of multi-turn, tool-using LLM reasoning agents. Tun...","url":"https://www.aioga.com/en/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:25:14.143Z"},"ja":{"title":"Google、Tunixを発表：JAXに基づく高スループットのエージェント後訓練ライブラリ","summary":"Google は Tunix を発表しました。これは JAX に基づくネイティブな後処理トレーニングライブラリであり、多段階でツールを使用する LLM 推論エージェントのトレーニングにおける TPU のアイドルボトルネックを解消することを目的としています。Tunix は高並行非同期ロールアウトと分離された生産者-消費者パイプラインを通じてハードウェアのスループットを最大化し、トレーナーが常にデータを取得できるようにします。このライブラリはプラグアンドプレイの抽象化と継続的なマクロレベルのパフォーマンス分析を提供し、カスタム環境の統合を容易にします。","category":"製品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google、Tunixを発表：JAXに基づく高スループットのエージェント後訓練ライブラリ - Aioga AIニュース","description":"Google は Tunix を発表しました。これは JAX に基づくネイティブな後処理トレーニングライブラリであり、多段階でツールを使用する LLM 推論エージェントのトレーニングにおける TPU のアイドルボトルネックを解消することを目的としています。Tunix は高並行非同期ロールアウトと分離された生産者-消費者パイプラインを通じてハードウェアのスルー...","url":"https://www.aioga.com/ja/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:25:34.175Z"},"ko":{"title":"구글, JAX 기반 고속 처리 지능 에이전트 사후 학습 라이브러리 Tunix 출시","summary":"구글은 JAX 기반의 네이티브 후처리 라이브러리 Tunix를 발표했으며, 이는 다중 라운드 및 도구 사용 LLM 추론 에이전트 훈련에서 TPU 유휴 병목 현상을 제거하는 것을 목표로 합니다. Tunix는 고동시 비동기 롤아웃과 분리된 생산자-소비자 파이프라인을 통해 하드웨어 처리량을 최대화하고, 트레이너가 지속적으로 데이터를 받을 수 있도록 보장합니다. 이 라이브러리는 플러그인 방식의 추상화와 지속적인 거시적 성능 분석을 제공하여 맞춤형 환경 통합을 용이하게 합니다.","category":"제품 업데이트","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"구글, JAX 기반 고속 처리 지능 에이전트 사후 학습 라이브러리 Tunix 출시 - Aioga AI 뉴스","description":"구글은 JAX 기반의 네이티브 후처리 라이브러리 Tunix를 발표했으며, 이는 다중 라운드 및 도구 사용 LLM 추론 에이전트 훈련에서 TPU 유휴 병목 현상을 제거하는 것을 목표로 합니다. Tunix는 고동시 비동기 롤아웃과 분리된 생산자-소비자 파이프라인을 통해 하드웨어 처리량을 최대화하고, 트레이너가 지속적으로 데이...","url":"https://www.aioga.com/ko/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:26:18.916Z"},"es":{"title":"Google lanza Tunix: biblioteca de postentrenamiento de agentes inteligentes de alto rendimiento basada en JAX","summary":"Google lanzó Tunix, una biblioteca nativa de postentrenamiento basada en JAX, diseñada para eliminar los cuellos de botella de inactividad de TPU en el entrenamiento de agentes de inferencia LLM de múltiples rondas que utilizan herramientas. Tunix maximiza el rendimiento del hardware mediante rollouts asincrónicos de alta concurrencia y una canalización desacoplada de productor-consumidor, asegurando que el entrenador reciba datos de manera continua. La biblioteca ofrece abstracciones plug-and-play y análisis de rendimiento a gran escala en tiempo real, facilitando la integración de entornos personalizados.","category":"Productos","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google lanza Tunix: biblioteca de postentrenamiento de agentes inteligentes de alto rendimiento basada en JAX - Aioga Noticias de IA","description":"Google lanzó Tunix, una biblioteca nativa de postentrenamiento basada en JAX, diseñada para eliminar los cuellos de botella de inactividad de TPU en el entrenamiento de agentes de...","url":"https://www.aioga.com/es/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:26:15.637Z"},"fr":{"title":"Google lance Tunix : une bibliothèque de post-formation intelligente à haut débit basée sur JAX","summary":"Google a publié Tunix, une bibliothèque native post-entraînement basée sur JAX, visant à éliminer le goulot d'étranglement des TPU inactifs lors de la formation d'agents de raisonnement LLM utilisant plusieurs tours et des outils. Tunix maximise le débit matériel grâce à un déploiement asynchrone à haute concurrence et à une pipeline producteur-consommateur découplée, garantissant que l'entraîneur reçoit constamment des données. La bibliothèque offre des abstractions plug-and-play et une analyse continue des performances à grande échelle, facilitant l'intégration d'environnements personnalisés.","category":"Produits","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google lance Tunix : une bibliothèque de post-formation intelligente à haut débit basée sur JAX - Aioga Actualités IA","description":"Google a publié Tunix, une bibliothèque native post-entraînement basée sur JAX, visant à éliminer le goulot d'étranglement des TPU inactifs lors de la formation d'agents de raisonn...","url":"https://www.aioga.com/fr/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:27:08.479Z"},"de":{"title":"Google stellt Tunix vor: Eine auf JAX basierende Bibliothek für das Training von intelligenten Agenten mit hoher Durchsatzrate","summary":"Google veröffentlicht Tunix, eine auf JAX basierende native Bibliothek für Post-Training, die darauf abzielt, Flaschenhälse durch ungenutzte TPUs beim Training von LLM-Inferenzagenten mit mehreren Runden und Werkzeugnutzung zu beseitigen. Tunix maximiert die Hardware-Durchsatzleistung durch hochgradig parallele asynchrone Rollouts und entkoppelte Produzenten-Konsumenten-Pipelines und stellt sicher, dass der Trainer kontinuierlich Daten erhält. Die Bibliothek bietet sofort einsatzbereite Abstraktionen und kontinuierliche makroskopische Leistungsanalyse, was die Integration benutzerdefinierter Umgebungen erleichtert.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google stellt Tunix vor: Eine auf JAX basierende Bibliothek für das Training von intelligenten Agenten mit hoher Durchsatzrate - Aioga KI-News","description":"Google veröffentlicht Tunix, eine auf JAX basierende native Bibliothek für Post-Training, die darauf abzielt, Flaschenhälse durch ungenutzte TPUs beim Training von LLM-Inferenzagen...","url":"https://www.aioga.com/de/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:27:00.487Z"},"pt-BR":{"title":"Google lança Tunix: biblioteca de pós-treinamento para agentes inteligentes de alta capacidade baseada em JAX","summary":"O Google lançou o Tunix, uma biblioteca nativa de pós-treinamento baseada em JAX, projetada para eliminar os gargalos de ociosidade de TPU no treinamento de agentes de inferência LLM multi-turno que utilizam ferramentas. O Tunix maximiza a taxa de transferência do hardware por meio de rollout assíncrono de alta concorrência e pipelines de produtor-consumidor desacoplados, garantindo que o treinador receba dados continuamente. A biblioteca oferece abstrações plug-and-play e análise de desempenho contínua em nível macro, facilitando a integração de ambientes personalizados.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google lança Tunix: biblioteca de pós-treinamento para agentes inteligentes de alta capacidade baseada em JAX - Aioga Notícias de IA","description":"O Google lançou o Tunix, uma biblioteca nativa de pós-treinamento baseada em JAX, projetada para eliminar os gargalos de ociosidade de TPU no treinamento de agentes de inferência L...","url":"https://www.aioga.com/pt-BR/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:27:50.340Z"},"ru":{"title":"Google выпустила Tunix: библиотеку дообучения интеллектуальных агентов с высокой пропускной способностью на основе JAX","summary":"Google выпустила Tunix, нативную библиотеку посттренировки на основе JAX, предназначенную для устранения простаивания TPU при обучении интеллектуальных агентов LLM с многократными раундами и использованием инструментов. Tunix максимизирует пропускную способность оборудования за счет высокоэффективного асинхронного rollout и разъединенной конвейерной схемы производитель-потребитель, обеспечивая постоянное получение данных тренажером. Библиотека предоставляет готовые к использованию абстракции и непрерывный макроуровневый анализ производительности, упрощая интеграцию пользовательских сред.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google выпустила Tunix: библиотеку дообучения интеллектуальных агентов с высокой пропускной способностью на основе JAX - Aioga Новости ИИ","description":"Google выпустила Tunix, нативную библиотеку посттренировки на основе JAX, предназначенную для устранения простаивания TPU при обучении интеллектуальных агентов LLM с многократными...","url":"https://www.aioga.com/ru/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:27:56.105Z"},"ar":{"title":"جوجل تطلق Tunix: مكتبة تدريب لاحق لوكلاء ذكيين عالية الإنتاجية تعتمد على JAX","summary":"أصدرت Google مكتبة Tunix، وهي مكتبة بعد التدريب أصلية مبنية على JAX، تهدف إلى القضاء على عنق الزجاجة الناتج عن عدم استخدام TPU في تدريب وكلاء الذكاء الاصطناعي القادرين على الاستدلال متعدد الجولات واستخدام الأدوات. تعمل Tunix على تعظيم كفاءة الأجهزة من خلال النشر المتزامن عالي التوازي وخط أنابيب المنتج-المستهلك المفصول، مما يضمن استمرار الحصول على البيانات من قبل المدرب. توفر المكتبة تجريدات قابلة للتوصيل والفصل وتحليل الأداء المستمر على المستوى الكلي، ما يسهل دمج البيئات المخصصة.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"جوجل تطلق Tunix: مكتبة تدريب لاحق لوكلاء ذكيين عالية الإنتاجية تعتمد على JAX - Aioga أخبار الذكاء الاصطناعي","description":"أصدرت Google مكتبة Tunix، وهي مكتبة بعد التدريب أصلية مبنية على JAX، تهدف إلى القضاء على عنق الزجاجة الناتج عن عدم استخدام TPU في تدريب وكلاء الذكاء الاصطناعي القادرين على الاستدلا...","url":"https://www.aioga.com/ar/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:28:43.219Z"},"hi":{"title":"Google ने Tunix लॉन्च किया: JAX पर आधारित उच्च थ्रूपुट इंटेलिजेंट एजेंट पोस्ट-ट्रेनिंग लाइब्रेरी","summary":"Google ने Tunix जारी किया, एक JAX आधारित मूल पोस्ट-ट्रेनिंग लाइब्रेरी, जिसका उद्देश्य मल्टी-राउंड और टूल-यूजिंग LLM इंफरेंस एजेंट प्रशिक्षण में TPU निष्क्रियता की समस्या को समाप्त करना है। Tunix उच्च समवर्ती असिंक्रोनस रोलआउट और अलग-प्लगिन प्रोड्यूसर-कंज्यूमर पाइपलाइन के माध्यम से हार्डवेयर थ्रूपुट को अधिकतम करता है, ताकि प्रशिक्षणकर्ता लगातार डेटा प्राप्त कर सके। यह लाइब्रेरी प्लग-एंड-प्ले एब्स्ट्रैक्शन और निरंतर मैक्रो-लेवल प्रदर्शन विश्लेषण प्रदान करती है, जिससे कस्टम वातावरण को एकीकृत करना आसान हो जाता है।","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google ने Tunix लॉन्च किया: JAX पर आधारित उच्च थ्रूपुट इंटेलिजेंट एजेंट पोस्ट-ट्रेनिंग लाइब्रेरी - Aioga AI समाचार","description":"Google ने Tunix जारी किया, एक JAX आधारित मूल पोस्ट-ट्रेनिंग लाइब्रेरी, जिसका उद्देश्य मल्टी-राउंड और टूल-यूजिंग LLM इंफरेंस एजेंट प्रशिक्षण में TPU निष्क्रियता की समस्या को समाप्त...","url":"https://www.aioga.com/hi/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:28:53.531Z"},"it":{"title":"Google lancia Tunix: libreria di addestramento post-agente ad alta capacità basata su JAX","summary":"Google ha rilasciato Tunix, una libreria nativa post-allenamento basata su JAX, progettata per eliminare il collo di bottiglia dell'inattività delle TPU nell'addestramento di agenti LLM per inferenze multi-turno e con utilizzo di strumenti. Tunix massimizza il throughput hardware tramite rollout asincroni ad alta concorrenza e una pipeline produttore-consumatore disaccoppiata, garantendo che l'allenatore riceva costantemente dati. La libreria offre astrazioni plug-and-play e analisi delle prestazioni a livello macro continuo, facilitando l'integrazione di ambienti personalizzati.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google lancia Tunix: libreria di addestramento post-agente ad alta capacità basata su JAX - Aioga Notizie IA","description":"Google ha rilasciato Tunix, una libreria nativa post-allenamento basata su JAX, progettata per eliminare il collo di bottiglia dell'inattività delle TPU nell'addestramento di agent...","url":"https://www.aioga.com/it/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:29:48.843Z"},"nl":{"title":"Google lanceert Tunix: een op JAX gebaseerde hoogdoorvoersbibliotheek voor training van intelligente agenten","summary":"Google heeft Tunix uitgebracht, een op JAX gebaseerde native library voor post-training, bedoeld om de TPU-idle bottleneck bij het trainen van LLM-agenten voor multi-turn en tool-gebruik te elimineren. Tunix maximaliseert de hardwaredoorvoer via high-concurrency asynchrone rollouts en een ontkoppelde producer-consumer pijplijn, waardoor de trainer continu data ontvangt. De library biedt plug-and-play abstracties en voortdurende macro-level prestatietesten, wat integratie van aangepaste omgevingen vergemakkelijkt.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google lanceert Tunix: een op JAX gebaseerde hoogdoorvoersbibliotheek voor training van intelligente agenten - Aioga AI-nieuws","description":"Google heeft Tunix uitgebracht, een op JAX gebaseerde native library voor post-training, bedoeld om de TPU-idle bottleneck bij het trainen van LLM-agenten voor multi-turn en tool-g...","url":"https://www.aioga.com/nl/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:29:38.267Z"},"tr":{"title":"Google Tunix'i Tanıttı: JAX Tabanlı Yüksek Verimli Ajan Sonrası Eğitim Kütüphanesi","summary":"Google, çok turlu, araç tabanlı LLM çıkarım ajanı eğitimindeki TPU boş darboğazını ortadan kaldırmak için tasarlanmış, JAX tabanlı yerel bir eğitim sonrası kütüphanesi olan Tunix'i piyasaya sürdü. Tunix, yüksek eşzamanlı asenkron dağıtımlar ve ayırılmış üretici-tüketici boru hatları aracılığıyla donanım verimliliğini en üst düzeye çıkarır; böylece eğitmenler sürekli veri alır. Kütüphane, takıp kullan soyutlama ve sürekli makro seviyesinde performans analizi sunarak özel ortamlarla entegrasyonu kolaylaştırır.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google Tunix'i Tanıttı: JAX Tabanlı Yüksek Verimli Ajan Sonrası Eğitim Kütüphanesi - Aioga AI Haberleri","description":"Google, çok turlu, araç tabanlı LLM çıkarım ajanı eğitimindeki TPU boş darboğazını ortadan kaldırmak için tasarlanmış, JAX tabanlı yerel bir eğitim sonrası kütüphanesi olan Tunix'i...","url":"https://www.aioga.com/tr/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:30:44.784Z"},"vi":{"title":"Google ra mắt Tunix: Thư viện huấn luyện sau cho tác nhân thông minh dựa trên JAX với tốc độ xử lý cao","summary":"Google phát hành Tunix, một thư viện huấn luyện hậu xử lý nguyên bản dựa trên JAX, nhằm loại bỏ nút thắt TPU nhàn rỗi trong việc huấn luyện các tác nhân suy luận LLM sử dụng nhiều vòng và công cụ. Tunix tối đa hóa thông lượng phần cứng thông qua rollout bất đồng bộ song song cao và pipeline nhà sản xuất-người tiêu dùng tách rời, đảm bảo bộ huấn luyện liên tục nhận dữ liệu. Thư viện này cung cấp các trừu tượng cắm và chạy cũng như phân tích hiệu suất liên tục ở cấp độ vĩ mô, thuận tiện cho việc tích hợp môi trường tùy chỉnh.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google ra mắt Tunix: Thư viện huấn luyện sau cho tác nhân thông minh dựa trên JAX với tốc độ xử lý cao - Tin tức AI Aioga","description":"Google phát hành Tunix, một thư viện huấn luyện hậu xử lý nguyên bản dựa trên JAX, nhằm loại bỏ nút thắt TPU nhàn rỗi trong việc huấn luyện các tác nhân suy luận LLM sử dụng nhiều...","url":"https://www.aioga.com/vi/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:30:41.465Z"},"id":{"title":"Google meluncurkan Tunix: Perpustakaan pelatihan lanjutan agen cerdas berbasis JAX dengan throughput tinggi","summary":"Google merilis Tunix, sebuah perpustakaan pasca-pelatihan asli berbasis JAX, yang bertujuan untuk menghilangkan hambatan TPU yang menganggur dalam pelatihan agen inferensi LLM multi-langkah yang menggunakan alat. Tunix memaksimalkan throughput perangkat keras melalui rollout asinkron berkonkurensi tinggi dan jalur produksi-konsumen yang terpisah, memastikan pelatih terus menerima data. Perpustakaan ini menyediakan abstraksi plug-and-play dan analisis kinerja makro berkelanjutan, memudahkan integrasi dengan lingkungan kustom.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google meluncurkan Tunix: Perpustakaan pelatihan lanjutan agen cerdas berbasis JAX dengan throughput tinggi - Berita AI Aioga","description":"Google merilis Tunix, sebuah perpustakaan pasca-pelatihan asli berbasis JAX, yang bertujuan untuk menghilangkan hambatan TPU yang menganggur dalam pelatihan agen inferensi LLM mult...","url":"https://www.aioga.com/id/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:31:30.130Z"},"th":{"title":"Google เปิดตัว Tunix: ไลบรารีการฝึกหลังสำหรับเอเยนต์อัจฉริยะที่มีประสิทธิภาพสูงบนพื้นฐาน JAX","summary":"Google เปิดตัว Tunix ซึ่งเป็นไลบรารีหลังการฝึกแบบเนทีฟที่สร้างบน JAX โดยมีวัตถุประสงค์เพื่อกำจัดปัญหาคอขวดของ TPU ในการฝึกตัวแทนสังเคราะห์ LLM ที่ใช้เครื่องมือหลายรอบ Tunix เพิ่มประสิทธิภาพการใช้งานอุปกรณ์โดยใช้ rollout แบบอะซิงโครนัสความพร้อมสูงและสายการผลิตแบบผู้ผลิต-ผู้บริโภคแยกส่วน เพื่อให้แน่ใจว่าตัวฝึกได้รับข้อมูลอย่างต่อเนื่อง ไลบรารีนี้มีการใช้ abstraction แบบ plug-and-play และการวิเคราะห์ประสิทธิภาพระดับมหภาคอย่างต่อเนื่อง ทำให้ง่ายต่อการรวมเข้ากับสภาพแวดล้อมที่กำหนดเอง","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google เปิดตัว Tunix: ไลบรารีการฝึกหลังสำหรับเอเยนต์อัจฉริยะที่มีประสิทธิภาพสูงบนพื้นฐาน JAX - ข่าว AI Aioga","description":"Google เปิดตัว Tunix ซึ่งเป็นไลบรารีหลังการฝึกแบบเนทีฟที่สร้างบน JAX โดยมีวัตถุประสงค์เพื่อกำจัดปัญหาคอขวดของ TPU ในการฝึกตัวแทนสังเคราะห์ LLM ที่ใช้เครื่องมือหลายรอบ Tunix เพิ่มปร...","url":"https://www.aioga.com/th/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:31:37.260Z"},"pl":{"title":"Google wprowadza Tunix: wysokoprzepustowa biblioteka do treningu agentów inteligentnych oparta na JAX","summary":"Google ogłosiło Tunix, natywną bibliotekę do post-treningu opartą na JAX, mającą na celu wyeliminowanie wąskiego gardła nieaktywności TPU w treningu inteligentnych agentów LLM wielokrotnej interakcji z użyciem narzędzi. Tunix maksymalizuje przepustowość sprzętu poprzez równoległe asynchroniczne rollouty i oddzielony pipeline producent-konsument, zapewniając, że trener ciągle otrzymuje dane. Biblioteka oferuje abstrakcje typu plug-and-play oraz ciągłą analizę wydajności na makro poziomie, ułatwiając integrację niestandardowych środowisk.","category":"产品更新","source":"Google Developers Blog（RSS）","aggregationSource":"Google Developers Blog（RSS）","pageTitle":"Google wprowadza Tunix: wysokoprzepustowa biblioteka do treningu agentów inteligentnych oparta na JAX - Aioga Wiadomości AI","description":"Google ogłosiło Tunix, natywną bibliotekę do post-treningu opartą na JAX, mającą na celu wyeliminowanie wąskiego gardła nieaktywności TPU w treningu inteligentnych agentów LLM wiel...","url":"https://www.aioga.com/pl/news/cmruuc0d80006bii0tzwael16/","contentTranslated":true,"sourceHash":"9bbd089dcd250ddb","translatedAt":"2026-07-22T17:32:23.427Z"}}}}