Scale LLM and MoE training across NVIDIA GPUs with 5D parallelism.
npx skills add https://github.com/hung-phan/ml-skills --skill nemo
Please help me install this Agent Skill. Skill: nemo Source: https://github.com/hung-phan/ml-skills/tree/main/skills/ml-review/references/ml-libraries/nemo Command: npx skills add https://github.com/hung-phan/ml-skills --skill nemo
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.
Use the nemo skill to train a large LLM model on a multi-node GPU cluster.
High-intent search queries and answers about installing and using this skill.
To scale LLM and MoE training across multiple NVIDIA GPUs, you can use this Skill to apply 5D parallelism and MoE Parallel Folding across multi-node clusters. It leverages Megatron-Core to optimize distributed training workloads natively within the NVIDIA stack.
Yes, this Skill provides verified, production-ready training configurations for Llama, Qwen, and DeepSeek-V3 models. These built-in configurations ensure proper setup for pretraining and post-training across multi-node GPU clusters.
The best way to deploy trained models to NIM or TensorRT-LLM is using this Skill's built-in deployment path. It streamlines the transition from multi-node GPU training directly to inference using TensorRT-LLM, vLLM, and NIM frameworks.
Yes, you can use FP8 and MXFP8 precision for large language model training. The Skill natively supports these precision formats through the transformer-engine dependency to optimize memory and compute efficiency during multi-node training.
Yes, you need nvidia-docker and Megatron-Core dependencies to run multi-node LLM training. These components are required to manage the NVIDIA container environment and execute the 5D parallelism techniques used for scaling.
MoE Parallel Folding is used when training mixture-of-experts models on GPU clusters to optimize expert distribution and reduce communication overhead. This technique enables efficient scaling of MoE architectures across multiple nodes using the NVIDIA stack.
NVIDIA-BioNeMoFinetune pretrained KERMT encoder models on labeled CSV datasets.
ancolemanDeploy LLM and ML models for production inference with vLLM, TensorRT-LLM, Ollama, and BentoML.
zeronx798Optimize transformer attention kernels for speed and GPU memory on NVIDIA hardware.
ovachieverProcess distributed ML data pipelines with Ray Data across CPU/GPU clusters.
