Skill UI

Orchestrates PyTorch-native TorchTitan 4D parallelism (FSDP2, TP, PP, CP) for large-scale LLM pretraining, scaling from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing to boot Llama 3.1, DeepSeek V3, or custom models.

Gastown is a robust multi-agent orchestration system designed for Claude Code. It enables the simultaneous execution of multiple AI agents on complex projects. It manages the entire development lifecycle, from setting up workspaces and tracking work items (beads) to managing agent lifecycles (polecats/crew) and merging final code (Refinery). Use it when coordinating complex, parallel AI tasks.

This advanced skill is a multi-agent research framework designed for executing deep, extensive research in parallel. It orchestrates a team of 10 specialized agents to gather, synthesize, and analyze information simultaneously. It is ideal for complex topics that require rapid, broad, and battle-tested investigation across multiple dimensions, significantly accelerating the knowledge discovery process.

This comprehensive guide helps developers safely migrate between different versions of the Kling AI video generation model (e.g., v1.x to v2.x). It details breaking changes, parameter differences, feature availability matrices (including native audio and camera control), and provides code examples for parallel A/B testing and robust rollback strategies, ensuring seamless API adoption.

This skill provides comprehensive guidance on training Mixture of Experts (MoE) models, including architectures like Mixtral and DeepSeek-V3. It details techniques for achieving large-scale model capacity with significantly reduced computational costs (e.g., 5x reduction vs. dense models). Topics covered include top-k routing, load balancing, expert parallelism, and integration with frameworks like DeepSpeed and HuggingFace for efficient, resource-constrained training.

This guide details advanced architectural patterns for building complex AI systems using multiple Language Model instances (agents). It addresses the single-agent context bottleneck by distributing work, enabling parallelism, and facilitating specialization across different subtasks. Key patterns include Supervisor/Orchestrator, Peer-to-Peer/Swarm, and Hierarchical designs, providing guidelines for context isolation and effective coordination protocols.

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

PufferLib is a high-performance library designed for advanced reinforcement learning simulations and training. It excels in parallel environment simulation, offering optimized vectorization and native support for multi-agent systems. Utilizing algorithms like PuffeRL (optimized PPO+LSTM), it achieves millions of steps per second, significantly accelerating research and deployment across single-agent and multi-agent environments.

Stable Baselines3 is a robust, PyTorch-based library providing production-ready implementations of core reinforcement learning algorithms (PPO, SAC, DQN, etc.). It facilitates the training of single-agent RL models, supports the creation of custom Gymnasium environments, and handles advanced parallel training via vectorized environments. It is ideal for rapid prototyping and rigorous RL experiments.

Coordinates a pre-task intelligence protocol that activates all relevant agents, synthesizes parallel analyses, estimates real execution time, maps probable risks, and produces a contingency-aware plan before executing the user’s request.

Performs a full GEO audit of any website by discovering business type, crawling prioritized pages, and delegating parallel subagents to evaluate AI visibility, platform readiness, technical infrastructure, EEAT content, and schema markup, then outputs a composite GEO score with prioritized actions.

Spawns N parallel AI subagents that compete on the same task using isolated git worktrees. A coordinator dispatches, monitors, evaluates results by metric or LLM judge, then merges the winning branch. Ideal for parallel optimization, content variations, refactors, and research exploration.