Nemotron-3-Super Fine-tuning with NeMo AutoModel#

This directory contains a Jupyter notebook and PEFT config for fine-tuning Nemotron-3-Super with NVIDIA NeMo AutoModel.

What’s included#

  • automodel_lora_cookbook.ipynb — End-to-end recipe: dataset creation, config setup, finetuning, and optional export.

  • base-peft-config-cookbook.yaml — PEFT (LoRA) config. Copy this into your AutoModel clone at examples/llm_finetuning/nemotron

  • text2sql.py — Dataset Target. Copy this into your AutoModel clone at nemo_automodel/components/datasets/llm

Prerequisites#

  • Multi-GPU machine (e.g. 8× H100 for Nemotron-Super; fewer GPUs for smaller models).

  • uv (for AutoModel env), or use pip/venv as in the AutoModel repo docs.

  • Hugging Face token for gated models (e.g. Nemotron-Super): run huggingface-cli login.

Setup#

1. Clone the AutoModel repo#

Clone NVIDIA NeMo AutoModel and set up its environment (this repo does not include AutoModel):

git clone https://github.com/NVIDIA-NeMo/Automodel.git <AUTOMODEL_DIR> 
cd <AUTOMODEL_DIR>
uv venv && source .venv/bin/activate
uv sync --frozen

Optional (e.g. for Nemotron-Super): pip install mamba-ssm --no-build-isolation Optional: pip install transformer-engine-torch==2.11.0

2. Download the base model#

Obtain the base model (e.g. from Hugging Face or NVIDIA NGC) and place it on disk. You will point the PEFT config and/or notebook to this path (model.pretrained_model_name_or_path).

For gated models (e.g. [nvidia/nemotron-super-*](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16)), ensure you are logged in: huggingface-cli login

3. Copy the PEFT config into AutoModel#

Copy this repo’s base-peft-config-cookbook.yaml into the AutoModel tree.

cp base-peft-config-cookbook.yaml <AUTOMODEL_DIR>/examples/llm_finetune/nemotron/

4. Set paths in the notebook#

Open automodel_lora_cookbook.ipynb and set in the first cell:

  • WORKSPACE_ROOT — Your workspace root (e.g. the directory that contains this repo and/or your data).

  • AUTOMODEL_DIR — Path to your AutoModel clone (where you ran uv sync).

  • DATASET_DIR — Directory for training.jsonl, validation.jsonl, and optionally test.jsonl. The notebook can generate these from BIRD-SQL, or you can place your own JSONL files there.

Running the notebook#

  1. Use the kernel that points to the AutoModel .venv (or install the same deps in your Jupyter env).

  2. Dataset creation (Section 2) writes training.jsonl, validation.jsonl, and optionally test.jsonl into DATASET_DIR. Adjust the config so those paths match what you set in the PEFT config.

  3. Finetuning (Section 3) runs torchrun with the cookbook config. Ensure N_DEVICES matches your GPU count.