TechPrufer
Get started
Coreweave
Coreweave logoCoreweave · Feature trail

Agentic Workflow Infrastructure

36 posts · Jan 2026 – Oct 2026

What this trail covers

CoreWeave enhances its AI infrastructure for scaling complex agentic workloads by integrating serverless reinforcement learning (RL), production inference at scale, fleet-wide observability for agents and infrastructure, and autonomous improvement to continually boost agent reliability using real-time data. This post details the integration of serverless RL for post-training LLMs, CoreWeave Inference for predictable production execution, W&B Weave for end-to-end observability and failure mode de. The post introduces CoreWeave Model Distillation for turning production signals into training data, a programmable training API for custom training loops, and RL Rollouts on CoreWeave Inference for seamless integration of inference into the training loop, enabling faster iteration and model improvement.

All Coreweave trails
  1. Closing the Inference to Post-Training Loop | CoreWeave BlogIntroduces CoreWeave Model Distillation to capture inference traffic and turn it into training data for relabeling and iterating on models. Details a programmable training API for Serverless Training, allowing users to write training loops in Python while CoreWeave manages the infrastructure, enabling mixed SFT and RL approaches. Enhances CoreWeave Inference with RL Rollouts built on NVIDIA Dynamo with vLLM support, allowing hot-loading of new weights onto live deployments without impacting in-flight requests, thus accelerating the inference-to-training loop for RL.
  2. CoreWeave Registry: Manage, Discover, and Deploy AI AssetsThis post introduces the CoreWeave Registry, a system of record for managing AI assets (models, agents, datasets, prompts). It adds capabilities for versioning, aliases, lineage tracking, permissions, and discovery. The registry aims to provide a foundation for reliable AI delivery by enabling teams to find, understand, and deploy assets with confidence. It supports a governed workflow from development to production, including managed imports from sources like Hugging Face, and integrates with CoreWeave Sandboxes for evaluation. The registry also facilitates shared discovery across the enterprise while maintaining access controls.
  3. What Is ARIA, and How Did We Build It?This post details the evolution of the ARIA agent, moving beyond a simple chat interface to a durable research process. Key technical contributions include: the concept of a persistent 'Turn' to store agent work and history, enabling investigations to survive browser sessions and recover from failures; a deliberate architectural separation of concerns (ARIA service, agent, environment, sandbox, driver) to improve failure analysis and modularity; and the use of Weave for end-to-end observability and offline evaluation of agent behavior against production-like conditions, using durable Turns as the basis for reproducible testing.
  4. How We Benchmark ARIA—and Why It Runs on WeaveThis post details the WBAF (W&B Agent Factory) framework for benchmarking ARIA, an agentic system. WBAF ensures production-aligned evaluations by using durable 'Turns' captured from the ARIA service, which include the prompt, history, agent configuration, and environment state. This approach contrasts with simple prompt-based benchmarking by preserving the full context of a production interaction. The framework uses Weave for connected observability and evaluation, allowing for the comparison of current and proposed agent configurations. Tasks are designed to resemble real jobs, preserving page context, conversational history, and external data dependencies. Evaluation uses a variety of scorers (Deterministic, Rubric, Trajectory, Reference) to check specific aspects of agent behavior, such as the existence of created artifacts, the quality of explanations, the evidence used, and adherence to truth. The benchmark suite is managed to remain relevant, focusing on the specific capabilities that have changed, with a broader regression suite available when needed.
  5. Why I Built Fugue: Controlled Experiments for Agent SystemsThis post introduces the Fugue framework for building and evaluating agent systems. It defines an agent system as comprising six parts: model, harness, tools, policy, memory, and runtime. Fugue enables controlled experiments by holding the workload and evaluator constant while systematically varying components of the agent system. The author details four engineering habits that improve individual agent runs: placing authoritative context near code, naming outcomes/constraints/verification/stop conditions, requiring receipts (evidence) for success, and testing tool exposure. The post highlights that the 'harness' (the execution path and run mechanics) significantly impacts agent performance, demonstrating a 23-percentage-point gap between two different harness implementations. This work shifts the focus from solely optimizing prompts to understanding and rigorously testing the entire agent system.
  7. CoreWeave Notebooks: Reactive Python Notebooks for AIIntroduces CoreWeave Notebooks, a new feature within CoreWeave Forge, leveraging marimo for reactive Python notebooks. These notebooks are designed for interactive AI development, enabling users to connect directly to project data and experiments. The post highlights their utility in debugging by identifying failure modes hidden in aggregate metrics (e.g., a driving model's performance degradation in specific weather conditions) and facilitating collaborative analysis through shared project storage. Future plans include publishing notebook views as persistent panels, hosting them as data apps, enabling direct model training, and integrating with ARIA for co-authoring.
  8. Accelerate AI research and iteration with CoreWeave ARIAIntroduces CoreWeave ARIA, an AI Research and Iteration Agent, designed to automate and streamline the AI development loop. ARIA analyzes experiments, generates reports with visualizations, proposes and launches new experiments based on project memory and connected code repositories (GitHub), and integrates with tools like Weights & Biases Models, Serverless RL, and Agent Lens. It offers visual analysis of data directly in chat and connects to external model/dataset repositories.
  9. CoreWeave Agent Lens: Automate Agent Observability and ImprovementIntroduces CoreWeave Agent Lens, a new product for automating agent observability and improvement. Agent Lens ingests OpenTelemetry tracing data to automatically surface semantic clusters of user intents and agent failures. It allows for failure detection based on the full production corpus, grouping similar failures into clusters with evidence and representative examples. It also clusters conversations by user intent to understand customer needs and capability gaps. Teams can create custom LLM judges from discovered failures to align evaluations with quality standards and convert production failures into test cases for candidate fix validation. This enables a continuous improvement loop by connecting production evidence to prompt updates, tool changes, model choices, or training runs, and supports integration with coding agents via Skills and MCP.
  10. Introducing CoreWeave Forge | CoreWeave BlogThis post introduces CoreWeave Forge, a platform designed to connect and accelerate the AI improvement lifecycle. It integrates several new and existing services to streamline the process of taking AI models and agents from iteration to production and back. Key contributions include: CoreWeave Agent Lens for end-to-end observability of agent execution, failure detection, and human oversight for curation; CoreWeave ARIA for AI-assisted research, coding, and iteration; CoreWeave Sandboxes for isolated CPU/GPU execution environments; Model Distillation for creating smaller, efficient models from larger ones; and CoreWeave Notebooks for collaborative experimentation. Forge also emphasizes the integration of CoreWeave Inference and Weights & Biases Models to support the entire AI development and deployment loop.
  11. NVIDIA Vera CPU: The First CPU for AI Agents | CoreWeave BlogIntroduces NVIDIA Vera CPUs to CoreWeave's bare-metal infrastructure, specifically targeting agentic AI workloads. Highlights over 3x faster agent sandbox startup times compared to x86 alternatives and the ability to run 11,000+ concurrent environments per rack due to Spatial Multithreading and DPU offloading. Details flexible capacity models (committed, serverless spillover, spot) for CPU usage and unified management with GPUs via CKS, SUNK, and Mission Control. Emphasizes Arm architecture consistency with GPU systems.
  12. Benchmarking GPT-6 Astra, Claude Fable 5.1, and GPT-5.6This post details a tutorial on benchmarking GPT-6 Astra, Claude Fable 5.1, and GPT-5.6 Sol using W&B Weave. It analyzes benchmark scores, token usage, response times, and estimated API costs across HumanEval coding tasks and other LLM evaluations. The post discusses benchmark saturation and the relationship between benchmark scores and cost per task, providing specific cost breakdowns for each model. It also highlights the use of agentic workflows for evaluating LLM performance in complex tasks.
  13. Getting started with the Agent Reinforcement Trainer (ART)This post introduces the Agent Reinforcement Trainer (ART), an open-source framework for applying reinforcement learning to LLM agents. It details the installation of ART and its dependencies (Python, pip, wandb, torchtune), and provides a concrete example of training a Qwen 2.5 model to play Tic-Tac-Toe. The example demonstrates defining a game environment, agent-environment interaction, reward assignment using an LLM-as-judge (RULER), and iterative model updates via GRPO. It highlights ART's integration with Hugging Face models and LoRA training for efficient fine-tuning.
  14. From Simulation-to-Reality for RoboticsThis post introduces domain randomization as a technique to bridge the sim-to-real gap in robotics. It details how randomizing parameters like gravity, floor friction, motor power, and sensor noise during training helps AI models generalize to real-world conditions. Additionally, it highlights the importance of tracking non-reward metrics such as action smoothness, average torque, and peak torque to detect hidden failure modes like motor overheating and mechanical wear, which are crucial for robust robotic systems.
  15. Agentic workflows in production: Mechanisms, costs and trade-offsThis post details the engineering challenges and solutions for operating agentic workflows in production. It introduces the concept of agentic workflows as distinct from RPA, traditional AI workflows, and chatbots, highlighting the importance of explicit goals, policies, tool contracts, state, budgets, and success criteria. The core mechanism described is a bounded control loop (plan, act, observe, reflect) with explicit state management, typed tool schemas, retry policies, loop limits, timeouts, cost budgets, and termination conditions. It also discusses the economic trade-offs and design patterns for cost optimization, such as context engineering and architecture choices, and emphasizes the need for observability and debugging capabilities for these systems.
  16. Introducing CoreWeave ARIA: AI Research and Iteration AgentIntroduces CoreWeave ARIA, an AI Research and Iteration Agent built into Weights & Biases. ARIA automates the research loop by analyzing experiments, forming hypotheses, writing configurations, launching experiments via W&B Launch, evaluating results, and drafting reports. It leverages deep platform knowledge of W&B for efficient querying and analysis across large projects and team boundaries. ARIA provides live, persistent dashboards to back up its analyses and can access training code, logs, metrics, artifacts, and checkpoints. The post also highlights ARIA's availability on mobile and outlines future roadmap items including shared agent conversations, extensible custom instructions, data source integration via MCP, and enterprise admin controls.
  17. CoreWeave Mission Control | Operational Intelligence for AIThis post introduces CoreWeave Mission Control MCP and Agent. MCP provides a hosted Model Context Protocol server, enabling MCP-compatible clients (like Claude Code, Cursor, Codex) to query metrics, logs, documentation, and resource state (clusters, nodes, SUNK, CAOIS) scoped to an organization via API token. This facilitates agent-driven investigations by bringing together disparate data sources. The Mission Control Agent integrates operational intelligence into the CoreWeave console, identifying performance bottlenecks in AI deployments (e.g., inference throughput drops due to KV cache pressure from longer contexts) and recommending specific optimizations like quantization (NVFP4, FP8) and model reshaping based on workload characterization, NVLink topology, and communication costs. It leverages hardware observability telemetry (chip, chassis, rack, network fabric, cooling) and internal performance data to provide actionable insights.
  18. Agentic Inference in Production: 4 Key Decisions | CoreWeaveThis post details four key infrastructure decisions for agentic inference in production: optimizing for Time-to-First-Token (TTFT) vs. End-to-End Completion, routing based on least-connections vs. session identity vs. prefix overlap, scaling on latency/throughput vs. concurrent requests with scale-to-zero vs. warm replica floor, and measuring cost per token vs. cost per task. It also provides six questions to help determine traffic shape and guide the choice of deployment model (how much of the serving stack to own).
  19. AI in Motorsports: How CoreWeave Helps JOTA Test Smarter | CoreWeaveThis post details the co-development of an AI-powered setup-recommendation tool for JOTA's motorsports engineering team. The tool uses active learning to identify the most informative car setups to test on a seven-post rig, aiming to maximize information gain within limited testing time. It models the car's performance response to setup changes based on historical data, highlighting areas of confidence and uncertainty. The tool surfaces potential setups, with engineers making the final decisions. A trackside application for damping adjustments is also under development.
  20. Multi-Plane Networking for Agentic AI: 5 Lessons | CoreWeaveThis post details CoreWeave's experience building a multi-plane network fabric for agentic AI. Key technical contributions include: 1. Implementing a non-blocking network fabric (1:1 uplink capacity at every layer) to prevent oversubscription and ensure low, predictable per-flow latency for agent chains alongside training traffic. This leverages NVIDIA Spectrum-X Ethernet, BlueField-3 DPUs, and ConnectX-9 SuperNICs. 2. Designing a multi-plane and multi-rail topology to scale the fabric without adding hops, allowing for a flat, two-tier design to reach large GPU counts while maintaining short, predictable GPU-to-GPU paths. 3. Utilizing hardware-accelerated load balancing on ConnectX-9 SuperNICs and adaptive routing on Spectrum-X switches to dynamically spread flows across available paths based on real-time congestion telemetry, avoiding static hashing issues. 4. Emphasizing resilience across five layers (topology, hardware, software, orchestration, and agent-level) to handle agent chain failures, as they lack clean restart boundaries like training checkpoints. 5. Tuning congestion control mechanisms (Priority Flow Control, Explicit Congestion Notification) against actual traffic patterns, particularly for NCCL collectives, to maintain a lossless fabric under demanding conditions.
  21. Video Captioning at Scale: 600 TB in 95 Minutes | CoreWeave BlogThis post details the successful deployment and scaling of a video captioning pipeline using Ray Data on CoreWeave's bare-metal Kubernetes infrastructure. It highlights the efficiency gains of using Ray Data over Ray Core for data processing and inference, achieving 3.7x more captions per GPU-hour. The post quantifies the performance of CoreWeave AI Object Storage (CAIOS) and its Local Object Transport Accelerator (LOTA) in handling large datasets (600 TB) with high throughput (40 GB/s writes, 110-120 GB/s reads) and low latency. It also showcases the rapid setup time (under 24 hours) from account signup to a production job running on 1,600 GPUs, emphasizing the synergy between Anyscale's workload orchestration and CoreWeave's managed infrastructure.
  22. Agentic Inference: Prefix Caching and Routing | CoreWeaveThis post introduces and details the concepts of prefix caching and prefix-aware routing for agentic inference. It explains how agentic prompts consist of a shared prefix and new content, and how reusing the KV cache for the prefix can significantly reduce prefill computation. The post emphasizes that prefix-aware routing is crucial to ensure that inference requests are directed to workers holding the relevant cached KV state, thereby avoiding cache misses. It outlines four key infrastructure considerations for production scale: routing to cached prefixes, preserving cache locality during scaling, balancing cache locality with load, and implementing robust observability for metrics like cache hit rate and TTFT. The post positions these techniques as essential for building infrastructure specifically designed for agentic workloads.
  23. CoreWeave's Stack for Physical AI | CoreWeave BlogThis post details CoreWeave's infrastructure stack for 'Physical AI' workloads, which involve real-world perception, decision-making, and action. It highlights how companies use CoreWeave for simulation-heavy training (MuJoCo, NVIDIA Isaac Sim, CARLA), multi-modal sensor fusion, and millisecond-latency inference. Specific examples include robotic manipulation, synthetic data generation, autonomous vehicle simulation, and reinforcement learning. The post emphasizes the importance of a robust infrastructure to support the 'observe, curate, improve, evaluate, run' flywheel for physical AI development, showcasing reference pipelines and blueprints available on CoreWeave.
  24. How the AI Loop Helps You Innovate Faster | CoreWeave BlogThis post introduces the 'AI Loop' concept, detailing a framework for continuous improvement of AI models and agents. It describes the integration of CoreWeave's inference and training platforms with Weights & Biases tools (Weave traces, Registry, Evaluations, Models, Automations) to create a feedback loop. Key components include running models/agents, observing their behavior, curating failures into datasets, evaluating improvements, and iterating. It highlights the distinction between agentic loops (prompt/skill improvements) and deeper model training loops (fine-tuning, RL on production data). The post also introduces CoreWeave ARIA as an AI Research and Iteration Agent that orchestrates this loop by analyzing runs, forming hypotheses, and launching experiments. It emphasizes the importance of observability across both the AI application and underlying infrastructure.
  25. Co-designed AI factories for production AI | CoreWeave BlogThis post details the co-design and validation process with NVIDIA for AI factories, focusing on the integration of GPUs, CPUs, networking, and storage to achieve high performance and efficiency for production AI workloads. It highlights the NVIDIA DSX Platform, Exemplar Cloud validation for both training and inference, and the extension of co-design principles to network and storage layers (NVIDIA Spectrum-X Ethernet, NVIDIA Quantum InfiniBand, CoreWeave AI Object Storage with LOTA). It also discusses the rapid bring-up of new hardware generations (GB300 NVL72) and the integration with CoreWeave's Kubernetes Service (CKS), SUNK, and observability tools.
  26. AI Agents Decode F1® Radio in Near Real Time | CoreWeave BlogThis post details the application of AI agents on CoreWeave Cloud to process Formula 1® radio communications in near real-time. It describes the challenge of managing 40 high-fidelity audio streams with a sub-five-second reaction time, requiring infrastructure capable of handling this data orchestration. The solution involves a platform with two modes: one for competitor channels (transcription, categorization, searchable interface) and another for internal team channels. A natural language chat interface allows engineers to query synthesized answers from live audio. The engineering challenge highlights the need for robust infrastructure, with 75 model iterations, 7 hours of hand-annotated data, and 3,000 labeled samples built for race conditions. The system achieved full transcription, diarization, and topic categorization within five seconds of capture. CoreWeave Cloud's AI platform, leveraging Weights & Biases Serverless Inference on CoreWeave Kubernetes Service and TensorRT-LLM, provides the necessary compute foundation. Fine-tuning was managed by Weights & Biases for experiment tracking and dataset versioning.
  27. Autonomous agents | CoreWeave BlogThis post details CoreWeave's approach to closing the loop between AI agent training and inference by integrating serverless reinforcement learning (RL), production inference at scale, and observability. It highlights how CoreWeave's Serverless RL enables iterative training on multi-turn agentic tasks, reducing training and inference loop times. The post also emphasizes the importance of CoreWeave Inference for reliable production execution and W&B Weave for monitoring agent behavior and detecting regressions. Finally, it introduces W&B Skills and MCP server for autonomous agent improvement.
  28. CoreWeave Sandboxes for Agentic AI | CoreWeave BlogIntroduces CoreWeave Sandboxes, an execution layer for agentic AI workloads (RL, tool use, evaluation). Details two deployment models: on CKS (Kubernetes pods with configurable network policies, resource limits) and serverless (Kata Containers by default for strong isolation, managed via Weights & Biases). Highlights integration with W&B for unified observability (lifecycle events, W&B Weave tracing) and secure secret injection. Emphasizes running sandboxes on existing compute (idle CPUs on GPU nodes) to reduce operational debt and cost.
  29. GLM 5.2 on CoreWeave Inference | CoreWeave BlogThis post announces the availability of GLM 5.2 on CoreWeave Inference, emphasizing its performance and price-performance for agentic workflows and software engineering tasks. It details the engineering effort involved in optimizing the inference stack for GLM 5.2, including model-level tuning, GPU and networking optimization, and serving runtime improvements, to achieve fast inference speeds and cost-efficiency. The post also highlights the benefits of using open-weight models like GLM 5.2 for production workflows and the ease of transitioning from evaluation to production on CoreWeave's managed platform.
  30. Infrastructure for Agentic Workflows | CoreWeave BlogThis post details how agentic workflows, characterized by multi-turn tool calls and bursty demand, introduce unique infrastructure challenges compared to traditional chatbots. It highlights the need for infrastructure that supports consistent latency across long dependency chains, scales rapidly to handle bursty inference demand, and provides predictable economics. The core argument is that the infrastructure's ability to manage these dynamic resource needs directly impacts the agent's responsiveness, reliability, and cost-effectiveness.
  31. Introducing NVIDIA HGX B300 on the Essential Cloud for AIThis post announces the general availability of NVIDIA HGX B300 on CoreWeave Cloud, specifically engineered for agentic AI workloads. It details how the HGX B300's enhanced interconnect speed (NVIDIA Quantum-X800 InfiniBand), BlueField-3 DPUs, and 800 Gbps ConnectX-8 SuperNICs improve NVFP4 inference performance and increase GPU memory capacity. Benchmarks demonstrate 3.42x faster token generation on Kimi K2.5 and 4.93x faster end-to-end request latency on DeepSeek-R1 compared to NVIDIA HGX H200. The post highlights the integration of HGX B300 with CoreWeave's existing AI infrastructure, including Mission Control for fleet management, CKS and SUNK for orchestration, and AI Object Storage for data access, emphasizing the platform's ability to accelerate agentic AI development through reduced friction, operational visibility, and high-throughput execution.
  32. Production AI Runs on Inference | CoreWeave BlogThis post shifts the focus to production-grade inference as the operational backbone of AI, particularly for agentic workflows. It highlights the challenges of scaling inference for production AI applications, emphasizing the need for predictable latency, reliability, and cost. The post introduces CoreWeave Inference with distinct deployment paths (Serverless and Dedicated) to address these needs, moving beyond just model deployment to operationalizing AI at scale. It discusses the evolution of inference demands from simple request-response to continuous reasoning and tool execution, and the importance of infrastructure, cost predictability, and control for mature workloads.
  33. Why Inference Is the Defining Layer of AI | CoreWeaveThis post shifts the focus from AI training to AI inference, highlighting its growing importance and the unique challenges posed by agentic AI workloads. It details how agentic inference compounds complexity through persistent GPU reservation, unpredictable I/O, high token generation, downstream dependencies, and concurrent agent coordination. The post also addresses systemic issues in production inference, such as cold starts, traffic spikes, observability gaps, and cost drift, and positions CoreWeave's full-stack infrastructure approach as a solution, emphasizing bare-metal GPUs, high-speed networking, AI Object Storage, and a flexible platform migration path from prototype to production.
  34. Why llm-d in CNCF Matters for Production Inference | CoreWeave BlogThis post announces the move of the llm-d project to the CNCF Sandbox, highlighting its significance for production inference as a cloud-native capability. It emphasizes llm-d's role in orchestrating and scaling distributed inference, addressing challenges in cost, performance, and portability for AI agents. The post details how llm-d provides an intelligent orchestration layer between serving frameworks and inference engines, working with Kubernetes-native components to manage complex distributed inference. CoreWeave's contribution is framed by its real-world experience in operating production inference, advocating for open, interoperable infrastructure.
  35. Achieve Full-Stack AI Observability: 4 Strategies | CoreWeaveThis post introduces strategies for achieving full-stack AI observability to address job failures and optimize AI infrastructure performance. It details four key areas: a cluster validation platform for pre-production readiness, regular active and passive health checks for runtime monitoring of nodes, networking, and storage, and emphasizes the benefits of bare-metal infrastructure for granular metric tracking. The post highlights how these observability measures contribute to reduced infrastructure management time, increased model training efficiency, higher resource utilization, cost savings, and faster time-to-market.
  36. Announcing distributed AI on CoreWeave with fully managed Ray on AnyscaleThis post announces the integration of Anyscale, a fully managed Ray platform, on CoreWeave Kubernetes Service (CKS) via BYOC. It details how this enables customers to run distributed AI workloads, including training, fine-tuning, inference, and multimodal data processing, directly within their CoreWeave accounts. Key benefits highlighted include enhanced performance and reliability through Anyscale's features such as efficient cluster validation, proactive health checking, world-class observability with granular metrics, and industry-leading goodput. The post also emphasizes Anyscale's developer tooling (Anyscale Workspaces), fully managed Ray clusters with autoscaling, RayTurbo for performance optimizations, and cost management features. It positions this integration as a shortcut to production for AI initiatives, particularly for RL workloads, by leveraging Anyscale's capabilities alongside CoreWeave's AI infrastructure.
  37. Breaking the Bottlenecks: Scaling AI Without Stalling | CoreWeave BlogThis post elaborates on the challenges of scaling AI infrastructure, including compute starvation, utilization inefficiencies, networking constraints, and storage drag. It critiques traditional scaling approaches as fragile and introduces strategies for building resilient AI infrastructure. Key strategies include designing for GPU elasticity, architecting for inevitable bursts, unifying observability, bringing compute closer to data, and designing for resilient multi-cloud/multi-AZ inference. It also discusses the benefits of purpose-built AI clouds and outlines future trends in AI infrastructure.
  38. CoreWeave Mission Control: CoreWeave’s AI Operating StandardThis post introduces CoreWeave Mission Control as an operating standard for AI at scale, focusing on reliability, transparency, and insights. Key new capabilities include Telemetry Relay for forwarding audit and observability signals to SIEMs, and GPU Straggler Detection for bottleneck analysis. It also previews the CoreWeave Mission Control Agent for conversational telemetry and remediation. The post highlights how Mission Control improves operational insight, enables proactive remediation, and contributes to higher goodput and model utilization by continuously evaluating cluster health, automating recoveries, and providing deep visibility into GPU, network, and storage behavior.
  39. Introducing Automated User Provisioning and SUNK User Provisioning | CoreWeave BlogThis post introduces Automated User Provisioning (AUP) and SUNK User Provisioning (SUP) to automate user provisioning for Slurm clusters. AUP federates identities from enterprise IdPs to CoreWeave IAM using SCIM, ensuring continuous synchronization of users and groups. SUP extends this automation to Slurm-on-Kubernetes (SUNK) clusters, automatically creating POSIX users, SSH keys, and Slurm users/accounts based on CoreWeave IAM data. It leverages NSSCache for resilient and fast authentication, distributing user and group data locally. This integration streamlines onboarding, reduces setup time from days to minutes, and enhances security by eliminating manual steps and stale credentials.
  40. Pretraining vs. Fine-Tuning vs. RAG: Choosing the Right AI ApproachThis post details the strategic decision-making process for choosing between AI model pretraining, fine-tuning, and RAG. It outlines the cost, time, control, and performance tradeoffs for each approach, providing real-world scenarios for financial services (pretraining), biotech research (fine-tuning), and enterprise IT (RAG). It also introduces a hybrid approach that combines these techniques, leveraging advancements like LoRA/QLoRA for fine-tuning and multi-hop retrieval/query decomposition for RAG. The post implicitly highlights the need for robust infrastructure to support these diverse AI development strategies, including GPU compute, networking, storage, and orchestration, which aligns with the capabilities of CoreWeave's AI infrastructure.
  41. Scaling Reinforcement Learning with torchforge on CoreWeave CloudThis post details the integration of torchforge, a PyTorch-scalable RL framework, with CoreWeave's Slurm on Kubernetes (SUNK) offering to enable large-scale Reinforcement Learning (RL) training and post-training workflows. It highlights a collaboration with Meta and Stanford's Scaling Intelligence Lab that successfully ran a dense coding model on 512 NVIDIA H100 GPUs using torchforge on CoreWeave Cloud. The post emphasizes how torchforge separates algorithmic logic from infrastructure complexity, and how SUNK's advanced scheduling and orchestration capabilities (handling over 32,000 GPUs) provide reliability, efficiency, and faster job startup for RL workloads.
  42. SkyPilot for Multi-Cloud Orchestration | CoreWeave BlogThis post details the integration of SkyPilot, an open-source framework for multi-cloud AI orchestration, with CoreWeave's infrastructure. Key contributions include: enabling one-line InfiniBand enablement for CoreWeave's networking tier, native integration with CoreWeave AI Object Storage (cw:// scheme), and autoscaling support for CoreWeave clusters. It also highlights the benefits of running SkyPilot on CoreWeave for AI teams, such as SLURM-like convenience on Kubernetes, unified control plane for multiple clusters, instant GPU provisioning, optimized price/performance, and reduced time-to-market due to fast networking and storage.
  43. Unlocking cross-cloud AI | CoreWeave BlogThis post introduces CoreWeave Interconnect, a private fiber-based network interconnect directly connecting CoreWeave Cloud to Google Cloud, leveraging Google Cloud's Partner Cross-Cloud Interconnect. It also details SUNK Anywhere, extending CoreWeave's SUNK (Slurm on Kubernetes) system to support GPU environments across clouds (including Google Cloud, AWS, Azure) or on-premises, enabling unified cross-cloud AI training. Additionally, it mentions LOTA Cross-Cloud for data transfer, though details are sparse. The post highlights the simplification of cross-cloud AI training and inference by reducing complexity, latency, and operational overhead compared to traditional third-party provider approaches.
  44. Powering Production-Ready Agentic AI with RAG | CoreWeave BlogThis post details the integration of vector databases (Milvus, Dragonfly, Pinecone) as a knowledge retrieval layer for RAG-based agentic AI on CoreWeave. It outlines two deployment strategies: running open-source vector databases directly on CoreWeave Kubernetes Service (CKS) for maximum control and data locality, and pairing CoreWeave with a managed service like Pinecone for a serverless option. The post describes the end-to-end workflow for agents using these databases, including generating embeddings on GPUs, querying vector databases, and feeding retrieved context into LLMs for reasoning and action. It also highlights how this enables specific agent patterns such as retrieval-augmented assistants, long-running agent workflows, and multimodal/recommendation-oriented agents.