nemo-mbridge-recipe-recommender

Megatron Bridge recipes are indexed based on your model name, GPU count, and training goal. The recommender matches your specifications to library and performance recipes, returning the best entry-point commands and dataset guidance.

Library recipes under src/megatron/bridge/recipes handle training via run_recipe.py, while performance recipes under scripts/performance measure throughput using dedicated perf scripts. The recommender distinguishes them to guide your entry points and defaults.

To avoid OOM, configure parallelism by ensuring TP divisibility of num_key_value_heads, staying within a node unless using interconnect, enabling SP when TP > 1, and adjusting micro-batch sizing based on DP implications.

Megatron Bridge supports PEFT and SFT training goals. The recommender indexes recipes based on your specified training objective, matching your model and hardware to appropriate library recipes for these tasks.

You need to enable sequence parallelism (SP) when tensor parallelism (TP) is greater than 1. The recommender enforces this critical configuration rule alongside TP divisibility of num_key_value_heads to ensure valid parallelism settings.

Yes, you can use Megatron Bridge recipes for GPU pretraining. The recommender indexes library recipes and advises on parallelism configurations like TP, PP, CP, and DP to fit your hardware, model size, and pretraining requirements.