Skill UI

Guides engineers through DeepSpeed distributed training and NVMe I/O optimizations, covering ZeRO, pipeline parallelism, mixed precision, and DeepNVMe handles for synchronous and asynchronous tensor storage.

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.

Guides automated distributed training setups with step-by-step practices, production-ready configs, and validation checks for ML training workflows.

HuggingFace Accelerate adds distributed support to any PyTorch script with just four lines, providing a unified API for DDP, DeepSpeed, FSDP, and Megatron, plus automatic device placement, mixed precision, and simple launch/config commands for fast prototyping.

High-performance RLHF framework that combines Ray, vLLM, and ZeRO-3 to accelerate PPO/GRPO/RLOO/DPO training for 7B-70B+ models on distributed GPU clusters, providing a single-stack workflow for reward models and policy optimization.

Guides adding PyTorch FSDP2’s fully_shard integration to training loops, covering correct distributed init, bottom-up sharding, mixed precision/offload policies, DTensor-aware optimizers, and distributed checkpointing so oversized models or DeviceMesh sharding goals work reliably.

PyTorch Lightning is a high-level framework built on PyTorch that significantly reduces boilerplate code in deep learning training loops. It automatically handles device management, distributed training (DDP, FSDP, DeepSpeed), mixed precision, and advanced callbacks. It allows researchers to scale models from local machines to supercomputers with minimal code changes, enabling clean, scalable, and best-practice deep learning implementations.

Ray Train orchestrates distributed PyTorch, TensorFlow, and HuggingFace workloads from laptops to clusters, automating GPU placement, fault tolerance, checkpointing, and hyperparameter sweeps so teams can scale massive model training with minimal code changes.

SkyPilot orchestrates ML training, batch jobs, and serving across 20+ clouds with automatic cost-aware region selection, spot-instance recovery, managed clusters, and GPU-aware distributed training, letting you optimize spend and avoid vendor lock-in.

torchforge is Meta's advanced PyTorch-native library designed for agentic Reinforcement Learning (RL). It fundamentally separates core RL algorithms from complex distributed infrastructure concerns. This allows researchers to rapidly prototype and experiment with novel RL methods (like GRPO) using clean, PyTorch-native abstractions, while automatically handling large-scale tasks like distributed training across thousands of GPUs using Monarch and TorchTitan.

This advanced workflow orchestrates complex, multi-node GPU clusters on Vast.ai. It is designed for large-scale distributed machine learning training, automatically handling spot instance preemption and resuming jobs from checkpoints. Furthermore, it provides comprehensive cost analysis by tracking billing history, allowing users to optimize GPU spending and manage cluster teardown efficiently.

Guides training large language models with Volcano Engine's verl library, covering PPO/GRPO workflows, distributed backends, multi-turn rollouts, reward configs, critic-enabled PPO, and production deployments using Ray workers.