nemo

What problem does it solve?

This Skill addresses the challenge of training large language models (LLMs) and mixture-of-experts (MoE) models at scale, specifically for users of the NVIDIA stack, requiring multi-node GPU clusters and advanced parallelism techniques.

Core Features & Use Cases

• Multi-node GPU Training: Optimizes training of LLMs and MoE models across multiple GPUs and nodes.
• Native NVIDIA Stack Support: Leverages Megatron-Core for 5D parallelism, FP8/MXFP8, and MoE Parallel Folding.
• Verified Configurations: Provides production-ready configurations for Llama, Qwen, and DeepSeek-V3 models.
• Deployment Path: Streamlines the process from training to deployment with TensorRT-LLM, vLLM, and NIM.
• Use Case: Ideal for scaling pretraining or post-training across many NVIDIA GPUs, training MoE models, and deploying to NIM/TRT-LLM.

Quick Start

Use the nemo skill to train a large LLM model on a multi-node GPU cluster.