nemo-automodel-recipe-development

What problem does it solve?

NeMo AutoModel recipe development is made accessible by providing a clear pattern for assembling and modifying training and evaluation recipes, including YAML structure, builder usage, and execution flow.

Core Features & Use Cases

• Guidance to create and modify Nemo AutoModel training and evaluation recipes across LLM, VLM, and diffusion domains.
• Explanation of the builder pattern (build_model, build_optimizer, build_dataloader, build_loss_fn, build_lr_scheduler, build_step_scheduler, build_checkpoint_config) and the required YAML anatomy.
• Use cases include adding new finetune recipes, updating existing configs, or debugging recipe execution and routing.

Quick Start

Copy the closest Nemo AutoModel recipe under nemo_automodel/recipes, update the model/dataset/dataloader, optimizer, and scheduler blocks, add or adjust a YAML example, then run automodel finetune llm -c config.yaml to validate locally.