Manage Model Entities for Customization

Before running a customization job, you need to set up a Model Entity that points to your base model checkpoint. This section covers creating the required FileSet and Model Entity.

Task Guides

Create a Model FileSet

Create a FileSet containing your base model checkpoint from Hugging Face, NGC, or local storage.

Create a Model Entity

Create a Model Entity that references your FileSet and enables customization.

Key Concepts

What is a FileSet?

A FileSet is a collection of files managed by the platform. For customization, you create a FileSet containing:

Model weights (.safetensors, .bin, or .nemo files)
Model configuration (config.json)
Tokenizer files (tokenizer.json, tokenizer_config.json, and so on)

FileSets can be populated from:

Hugging Face Hub - Download directly from HF repositories
NGC - Download from NVIDIA NGC catalogs
Local upload - Upload files from your local machine

What is a Model Entity?

A Model Entity is the platform’s representation of a model. It contains:

FileSet reference - Points to where the model files are stored
Model Spec - Auto-populated metadata about the model architecture
Adapters - LoRA or other adapters attached to this model (populated after training)
Base model link - For fine-tuned models, links back to the parent

Quick Start Example

Complete example of setting up a model for customization:

Hugging Face Token: If downloading from a gated Hugging Face repository, you will need to create a secret containing your Hugging Face API token first. Refer to Manage Secrets for instructions.

import time
from nemo_helix_plugin.client.client import NemoClient
from nemo_helix_plugin.client.errors import ConflictError
from nemo_helix_plugin.files.storage_config import HuggingfaceStorageConfig
from nemo_helix_plugin.files.types import CreateFilesetRequest
from nemo_helix_plugin.models.types import CreateModelEntityRequest
client = NemoClient.from_config()
# Step 1: Create FileSet from Hugging Face
try:
    fileset = client.files.create_fileset(
        body=CreateFilesetRequest(
            name="nemotron-3-5-lightning-30b-a3b",
            description="Nemotron 3.5 Lightning 30B A3B base model from Hugging Face",
            storage=HuggingfaceStorageConfig(
                repo_id="nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16",
                repo_type="model",
            ),
        ),
    ).data()
    print(f"Created FileSet: {fileset.name}")
except ConflictError:
    print("FileSet already exists, retrieving...")
    fileset = client.files.get_fileset(name="nemotron-3-5-lightning-30b-a3b").data()
# Step 2: Create Model Entity
try:
    model = client.models.create_model(
        body=CreateModelEntityRequest(
            name="nemotron-3-5-lightning-30b-a3b",
            fileset="default/nemotron-3-5-lightning-30b-a3b",  # Reference to the FileSet
            description="Nemotron 3.5 Lightning 30B A3B base model for customization",
        ),
    ).data()
    print(f"Created Model Entity: {model.name}")
except ConflictError:
    print("Model Entity already exists, retrieving...")
    model = client.models.get_model(name="nemotron-3-5-lightning-30b-a3b").data()
# Step 3: Wait for ModelSpec to be auto-populated
print("Waiting for model spec to be populated...")
while not model.spec:
    time.sleep(5)
    model = client.models.get_model(name="nemotron-3-5-lightning-30b-a3b").data()
print(f"Model ready!")
print(f" Architecture: {model.spec.family}")
print(f" Parameters: {model.spec.base_num_parameters:,}")
print(f" Layers: {model.spec.num_layers}")

These examples use the default workspace. If your current CLI context uses a different workspace, run nemo config set --workspace default first. To check, run nemo config view.

After the Model Entity is ready (has a spec), you can use it in customization jobs with model: "default/nemotron-3-5-lightning-30b-a3b".