YAML Configuration

NeMo AutoModel recipes are configured with YAML. Under the hood, YAML is parsed into a ConfigNode which:

• Translates common scalar strings into typed Python values (e.g., "10" → 10).
• Resolves _target_ (and *_fn) into Python callables/classes.
• Supports environment variable interpolation inside YAML strings.
• Tries to make config printing safe by preserving original placeholders (to avoid leaking secrets).

Load Model and Dataset Configs

Most recipes load the YAML using nemo_automodel.components.config.loader.load_yaml_config(), which returns a ConfigNode.

Within a ConfigNode:

• Nested dicts become nested ConfigNode objects.
• Lists are recursively wrapped.
• Scalars are translated with translate_value() when they are YAML strings.

Typed Scalar Translation (translate_value)

Only strings are translated. Examples:

• "123" → 123
• "3.14" → 3.14
• "true" / "false" → True / False
• "None" / "none" → None

YAML-native types (like step_size: 10 without quotes) are already typed by the YAML parser and remain unchanged.

Use _target_ for Instantiation

Any mapping containing a _target_ key can be instantiated using ConfigNode.instantiate():

1
model:
2
  _target_: nemo_automodel.NeMoAutoModelForCausalLM.from_pretrained
3
  pretrained_model_name_or_path: meta-llama/Llama-3.2-1B

There is also support for resolving callables from:

• Dotted paths: pkg.module.symbol
• Local file paths: /abs/path/to/file.py:symbol

Safety and Policy

By default, resolving targets is restricted:

• Imports are allowed from common safe prefixes (e.g. nemo_automodel, torch, transformers, …).
• Accessing private or dunder attributes is blocked by default.
• Loading out-of-tree user code can be enabled with NEMO_ENABLE_USER_MODULES=1 or by calling set_enable_user_modules(True).

Distributed Section (Strategy-Based)

The distributed: section is not instantiated using _target_. Recipes parse it with a fixed schema: use strategy: fsdp2, strategy: ddp, or strategy: megatron_fsdp, plus optional parallelism sizes (dp_size, tp_size, pp_size, etc.) and strategy-specific options. When pipeline parallelism is enabled (pp_size > 1), add a pipeline: subsection with options such as pp_schedule, pp_microbatch_size, and layers_per_stage. See the Pipeline Parallelism with AutoPipeline guide and recipe example configs for full examples.

Interpolate Environment Variables in YAML

NeMo AutoModel supports env var interpolation inside YAML string values.

Supported Forms

• Braced:
  • ${VAR}
  • ${VAR,default}
  • ${var.dot.var} (dots are treated as part of the env var name)
• Dollar:
  • $VAR
  • $var.dot.var
• Back-compat:
  • ${oc.env:VAR}
  • ${oc.env:VAR,default}

Interpolation Behavior

• Interpolation happens when values are wrapped into a ConfigNode.
• If a referenced env var is missing and no default is provided, config loading raises a KeyError.
• Defaults are supported only for braced forms using the first comma: ${VAR,default_value}.

Example (Databricks Delta)

1
dataset:
2
  _target_: nemo_automodel.components.datasets.llm.column_mapped_text_instruction_iterable_dataset.ColumnMappedTextInstructionIterableDataset
3
  path_or_dataset_id: delta://catalog.schema.training_data
4
  delta_storage_options:
5
    DATABRICKS_HOST: ${DATABRICKS_HOST}
6
    DATABRICKS_TOKEN: ${DATABRICKS_TOKEN}
7
    DATABRICKS_HTTP_PATH: ${DATABRICKS_HTTP_PATH}

Prevent Secret Leakage in Logs

When an env var placeholder is resolved, the config keeps the original placeholder in an internal ._orig_value field for safe printing:

• str(cfg) / repr(cfg) prints placeholders (e.g. ${DATABRICKS_TOKEN}), not resolved secrets.
• cfg.to_yaml_dict(use_orig_values=True, redact_sensitive=True) is the recommended way to produce a loggable YAML dict.

Configure Slurm

SLURM jobs are submitted with sbatch directly — no YAML section needed. Copy the reference script, edit CONFIG and cluster settings, then submit:

$
cp slurm.sub my_cluster.sub
$
vim my_cluster.sub
$
sbatch my_cluster.sub

All cluster-specific configuration (SBATCH directives, container image, mounts, secrets, environment variables) lives in your sbatch script. See Run on a Cluster for full examples.