Configure CoreWeave GPU clusters with node selectors, MIG profiles, and NCCL tuning.
npx skills add https://github.com/tomz/agent-skills --skill coreweave-gpu
Please help me install this Agent Skill. Skill: coreweave-gpu Source: https://github.com/tomz/agent-skills/tree/main/coreweave-gpu Command: npx skills add https://github.com/tomz/agent-skills --skill coreweave-gpu
CoreWeave GPU workloads are complex to plan, provision, and optimize across single-GPU to multi-node configurations, including MIG slices and NCCL tuning.
Use this skill to configure and optimize a CoreWeave GPU workload across a single node or multi-node cluster using the patterns in this Skill.
High-intent search queries and answers about installing and using this skill.
Configuring CoreWeave GPU workloads for multi-node training requires applying Kubernetes node selectors and multi-node patterns using PyTorch or Volcano. This enables efficient GPU allocation and scheduling across multiple nodes within your CoreWeave cluster.
MIG slicing in CoreWeave GPU clusters involves configuring mixed MIG and non-MIG pools using specific MIG profiles. This partitions a single GPU into isolated instances to maximize utilization and performance for smaller workloads.
The best way to tune NCCL for high bandwidth on CoreWeave is by setting specific NCCL environment variables across nodes. This optimization ensures robust performance tuning and efficient communication during multi-node GPU training operations.
Yes, you can use spot instances for CoreWeave GPU scheduling by implementing operational guardrails for spot versus on-demand pricing. This approach includes preemption handling and GPU resource safety measures to enable reliable workloads.
Maximizing GPU utilization and performance on CoreWeave requires applying GPU catalog guidance and targeted node selector strategies. These patterns ensure efficient GPU allocation and proper scheduling for Kubernetes-based workloads.
nist0Manage Kubernetes clusters with kubectl commands for workloads, logs, and resources.
jstzwjProvide collective operations for multi-GPU data communication in CUDA applications.
longkeyyAutomate content discovery across ArXiv, GitHub, and HuggingFace with semantic filtering and publishing.
BagelHoleManage GPU-backed Kubernetes clusters for AI inference and training.
