Managing Multiple Slurm Clusters with SkyPilot
If you have access to multiple Slurm clusters, managing them manually is painful. SkyPilot lets you use them as a single unified resource pool.
SkyPilot Blog
Guides, posts, and resources from SkyPilot.
Engineering
RL Is Bottlenecked by Inference. Scale It Independently.
Agentic RL spends most of every step on inference, not training. Scale inference in RL with SkyPilot Job Groups and slime.
Serving Kimi K3 on Your Own GPUs with SkyPilot
Kimi K3's weights are open. Here's how to serve the 2.8T-parameter model across your own GPU clusters with a single YAML.
Accelerating custom intelligence: Announcing SkyPilot Platform and our $20M funding
SkyPilot Endpoints: Production-Ready Inference on Every Cluster You Own
One YAML, one endpoint to run and manage inference on your own GPU fleet.
SkyPilot Sandboxes: Run Agent Code on Your Own Kubernetes, at Scale
SkyPilot runs untrusted, LLM-generated code in sandboxes on the Kubernetes clusters you already own. A single cluster sustains 50,000+ sandboxes, with multi-cluster support to go further. Individual launches in less than a second. All at up to one-tenth the cost of hosted alternatives, with your code and data never leaving your cloud.
RL Doesn't Work on Slurm
Online reinforcement learning for LLMs breaks Slurm's batch scheduling model. We'll discuss why, and what can be done about it.
Cache Me If You Can: Tuning Object Stores for AI
We ran hundreds of benchmarks to tune storage systems for distributed training so you don’t have to.
GPU Compass: Navigate the GPU Frontier Across 20+ Clouds & 2K+ Offerings
Introducing GPU Compass: One dashboard to browse, compare pricing, and launch across every GPU cloud.
SkyPilot Agent Skill: Let Agents Manage Your GPUs
With the SkyPilot Agent Skill, your AI coding agent can launch clusters, run training jobs and manage cloud resources across any infrastructure using natural language.
Research-Driven Agents: What Happens When Your Agent Reads Before It Codes
Coding agents working from code alone generate shallow hypotheses. Adding a research phase — arxiv papers, competing forks, other backends — produced 5 kernel fusions that made llama.cpp CPU inference 15% faster.
Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster
Karpathy's autoresearch runs one experiment at a time. We gave it access to our GPU infra and let it run experiments in parallel.
Subscribe to our monthly newsletter
Get the latest stories, tips, and insights delivered to your inbox.