nemo-automodel-recipe-development

Create and modify NeMo AutoModel training and evaluation recipes, including YAML structure, builders, and execution flow.

2,713stars315forksUpdated 7/29/2026

Data Science & ML #pytorch #configuration #machine-learning #nvidia #fine-tuning #yaml

Security Assessment

Safe(95/100)

Security Score95/100

About nemo-automodel-recipe-development

NeMo AutoModel Recipe Development is an NVIDIA-authored guidance skill for creating and modifying NeMo AutoModel training and evaluation recipes, covering YAML structure, builder functions, and execution flow. It helps engineers answer recipe questions with the smallest complete path to action: naming the relevant recipe file or YAML section, listing the builder functions or config keys involved, giving a minimal YAML or command example, and ending with a local validation command or CPU-compatible test.

The skill explains the recipe architecture end to end. It traces the execution flow from the CLI (automodel finetune llm) through app.py parsing, the recipe script's main entry, a Recipe class that inherits BaseRecipe and implements setup() and run_train_validation_loop(), and the builder pattern where components like model, optimizer, dataloader, loss, LR scheduler, step scheduler, and checkpoint config are each constructed by dedicated builder functions. It documents the strict infrastructure application order (PEFT, FP8, QAT, checkpoint restore, parameter freezing, sharding, device placement, torch.compile, context parallelism, and more), the meaning of _target_ fields as fully qualified Python callables with sibling keys as keyword arguments, nested CLI overrides, and key validation and checkpointing settings such as step_scheduler val and checkpoint intervals, validation_dataset, restore_from.path, and consolidated safetensors. It defines a routing boundary that keeps it focused on recipe construction rather than standalone distributed-strategy selection or cluster launcher setup. A skill card, benchmark report, and evaluation results accompany it, noting Apache-2.0 licensing, no required credentials, and a recommendation to review generated proposals before execution.

The intended users are developers and engineers building or debugging training, SFT, and evaluation recipes for NeMo AutoModel who need to add YAML fields, wire builders, register CLI routes, or write recipe tests.

FAQ

What does this skill help me build?

It helps create and modify NeMo AutoModel training and evaluation recipes, including YAML config structure, builder functions, CLI routing, and recipe-specific tests, and helps debug recipe construction or trainer issues.

How does a recipe execute?

From the CLI (automodel finetune llm -c config.yaml) through app.py parsing, into the recipe script's main, where a Recipe class subclassing BaseRecipe runs setup() to build components via builder functions and then run_train_validation_loop() to train.

What are _target_ fields?

A _target_ is the fully qualified Python callable to instantiate, and its sibling YAML keys become keyword arguments; you can override nested values from the CLI, for example --optimizer.lr.

What settings control validation and checkpointing?

step_scheduler.val_check_interval for validation cadence, step_scheduler.checkpoint_interval for save cadence, validation_dataset as the validation source, restore_from.path for resume, and consolidated safetensors as the default checkpoint format.

Does it require credentials or run training itself?

Per its skill card it requires no API key or external credential and is Apache-2.0 licensed. It is guidance that produces configuration and code suggestions, which should be reviewed before execution; it does not run training on your behalf.