Skill UI

Flash Attention optimizer delivers 2-4x speedup and 10-20x memory savings for transformer attention, ideal when training or running long-context models that hit GPU memory or latency limits; supports PyTorch native scaled_dot_product_attention, the flash-attn library, H100 FP8 precision, and sliding-window attention.

Provides production-ready Megatron-Core workflows for training 2B–462B parameter LLMs with tensor/pipeline/expert/context parallelism, maximizing GPU efficiency on H100 clusters.

A zero-config, serverless cloud platform designed for running intensive GPU workloads, including model training, fine-tuning, and inference. It eliminates the need for complex infrastructure setup (no SSH, no Docker) while supporting local-first development and auto-scaling for various high-performance GPUs (A100, H100). Ideal for data scientists and ML engineers requiring scalable and cost-efficient remote compute resources.

Modal is a powerful serverless cloud platform designed specifically for running AI and Machine Learning workloads. It allows users to deploy, serve, and scale complex models (like LLMs) without managing infrastructure. Key features include on-demand access to high-end GPUs (A100, H100), serverless function scaling, persistent storage, and the ability to define entire infrastructure using pure Python code, eliminating the need for separate YAML or Dockerfiles.

A comprehensive utility designed for managing CoreWeave cloud infrastructure migrations. It facilitates upgrading between GPU generations (e.g., A100 to H100), updating CUDA versions, and handling Kubernetes resource changes. The tool uses K8s APIs to audit deployments, detect deprecated instances, and execute controlled rollbacks.

This guide details how to set up and execute large-scale, distributed GPU training jobs on the CoreWeave platform. It covers both single-node multi-GPU setups and multi-node training utilizing PyTorch DDP and Kubernetes resources (A100/H100). This workflow is ideal for fine-tuning large language models (LLMs) or running complex deep learning models that require high-performance, shared-storage compute clusters.

This skill provides comprehensive guidance for deploying and managing GPU-accelerated AI inference services on CoreWeave Kubernetes (CKS). It covers best practices including containerization using NVIDIA CUDA base images, configuring specific GPU resource limits (A100/H100), setting up robust health checks, and executing controlled rolling updates. Ideal for managing multi-model inference and scaling demanding AI workloads in a cloud environment.

This guide demonstrates how to deploy GPU workloads on the CoreWeave platform using Kubernetes (kubectl). It covers deploying vLLM inference services for large language model (LLM) testing and running batch CUDA jobs for benchmarking. This is ideal for users running their first GPU workloads, testing inference pipelines, or verifying cluster access on bare-metal GPU nodes (A100, H100).

This SDK provides production-ready Python patterns for managing GPU workloads on CoreWeave. It includes helpers for defining specific GPU affinity, setting resource limits (A100, H100), and generating complete Kubernetes Deployment YAML. Use this library for building robust inference clients, managing GPU deployments programmatically, or automating complex AI workflows.