nemo-automodel-recipe-development

To create a NeMo AutoModel finetuning recipe, copy the closest existing recipe under nemo_automodel/recipes, update the model, dataset, dataloader, optimizer, and scheduler blocks, add a YAML example, then validate locally using automodel finetune llm -c config.yaml.

NeMo AutoModel recipes require builder functions including build_model, build_optimizer, build_dataloader, build_loss_fn, build_lr_scheduler, build_step_scheduler, and build_checkpoint_config to assemble the training and evaluation execution pipeline.

NeMo AutoModel YAML recipes require correct frontmatter with name and description fields, alongside properly structured blocks for model, dataset, dataloader, optimizer, and scheduler configurations to ensure proper CLI routing and execution.

Yes, NeMo AutoModel recipes support creating and modifying training and evaluation recipes across LLM, VLM, and diffusion domains, allowing you to configure builders and YAML structures for vision language model finetuning.

NeMo AutoModel recipe routing issues often stem from incorrect YAML frontmatter, specifically missing or malformed name and description fields, or errors in the builder pattern usage that disrupt the execution flow from CLI routing to the training loop.