dgx-spark-playbooks/nvidia/nemo-fine-tune/README.md

# Fine-tune with NeMo

> Use NVIDIA NeMo to fine-tune models locally

## Table of Contents

- [Overview](#overview)
- [Instructions](#instructions)
- [Troubleshooting](#troubleshooting)

---

## Overview

## Basic idea

This playbook guides you through setting up and using NVIDIA NeMo AutoModel for fine-tuning large language models and vision-language models on NVIDIA Spark devices. NeMo AutoModel provides GPU-accelerated, end-to-end training for Hugging Face models with native PyTorch support, enabling instant fine-tuning without conversion delays. The framework supports distributed training across single GPU to multi-node clusters, with optimized kernels and memory-efficient recipes specifically designed for ARM64 architecture and Blackwell GPU systems.

## What you'll accomplish

You'll establish a complete fine-tuning environment for large language models (1-70B parameters) and vision-language models using NeMo AutoModel on your NVIDIA Spark device. By the end, you'll have a working installation that supports parameter-efficient fine-tuning (PEFT), supervised fine-tuning (SFT), and distributed training capabilities with FP8 precision optimizations, all while maintaining compatibility with the Hugging Face ecosystem.

## What to know before starting

- Working in Linux terminal environments and SSH connections
- Basic understanding of Python virtual environments and package management
- Familiarity with GPU computing concepts and CUDA toolkit usage
- Experience with containerized workflows and Docker/Podman operations
- Understanding of machine learning model training concepts and fine-tuning workflows

## Prerequisites

- NVIDIA Spark device with Blackwell architecture GPU access
- CUDA toolkit 12.0+ installed and configured: `nvcc --version`
- Python 3.10+ environment available: `python3 --version`
- Minimum 32GB system RAM for efficient model loading and training
- Active internet connection for downloading models and packages
- Git installed for repository cloning: `git --version`
- SSH access to your NVIDIA Spark device configured

## Ancillary files

All necessary files for the playbook can be found [here on GitHub](https://github.com/NVIDIA-NeMo/Automodel)

## Time & risk

* **Duration:** 45-90 minutes for complete setup and initial model fine-tuning
* **Risks:** Model downloads can be large (several GB), ARM64 package compatibility issues may require troubleshooting, distributed training setup complexity increases with multi-node configurations
* **Rollback:** Virtual environments can be completely removed; no system-level changes are made to the host system beyond package installations.
* **Last Updated:** 01/15/2026
  * Fix qLoRA fine-tuning workflow

## Instructions

## Step 1. Verify system requirements

Check your NVIDIA Spark device meets the prerequisites for [NeMo AutoModel](https://github.com/NVIDIA-NeMo/Automodel) installation. This step runs on the host system to confirm CUDA toolkit availability and Python version compatibility.

```bash
## Verify CUDA installation
nvcc --version

## Check Python version (3.10+ required)
python3 --version

## Verify GPU accessibility
nvidia-smi

## Check available system memory
free -h

## Docker permission:
docker ps

## if there is permission issue, (e.g., permission denied while trying to connect to the Docker daemon socket), then do:
sudo usermod -aG docker $USER
newgrp docker
```

## Step 2. Configure Docker permissions

To easily manage containers without sudo, you must be in the `docker` group. If you choose to skip this step, you will need to run Docker commands with sudo.

Open a new terminal and test Docker access. In the terminal, run:

```bash
docker ps
```

If you see a permission denied error (something like permission denied while trying to connect to the Docker daemon socket), add your user to the docker group so that you don't need to run the command with sudo .

```bash
sudo usermod -aG docker $USER
newgrp docker
```

## Step 3. Get the container image with NeMo AutoModel

```bash
docker pull nvcr.io/nvidia/nemo-automodel:26.02
```

## Step 4. Launch Docker

Launch an interactive container with GPU access. The `--rm` flag ensures the container is removed when you exit.

```bash
docker run \
  --gpus all \
  --ulimit memlock=-1 \
  -it --ulimit stack=67108864 \
  --entrypoint /usr/bin/bash \
  --rm nvcr.io/nvidia/nemo-automodel:26.02
```

## Step 5. Explore available examples

Review the pre-configured training recipes available for different model types and training scenarios. These recipes provide optimized configurations for ARM64 and Blackwell architecture.

```bash
## Navigate to /opt/Automodel
cd /opt/Automodel

## List LLM fine-tuning examples
ls examples/llm_finetune/

## View example recipe configuration
cat examples/llm_finetune/finetune.py | head -20
```

## Step 6. Run sample fine-tuning
The following commands show how to perform full fine-tuning (SFT), parameter-efficient fine-tuning (PEFT) with LoRA and QLoRA.

First, export your HF_TOKEN so that gated models can be downloaded.

```bash
## Run basic LLM fine-tuning example
export HF_TOKEN=<your_huggingface_token>
```
> [!NOTE]
> Replace `<your_huggingface_token>` with your personal Hugging Face access token. A valid token is required to download any gated model.
>
> - Generate a token: [Hugging Face tokens](https://huggingface.co/settings/tokens), guide available [here](https://huggingface.co/docs/hub/en/security-tokens).
> - Request and receive access on each model's page (and accept license/terms) before attempting downloads.
>   - Llama-3.1-8B: [meta-llama/Llama-3.1-8B](https://huggingface.co/meta-llama/Llama-3.1-8B)
>   - Qwen3-8B: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B)
>   - Meta-Llama-3-70B: [meta-llama/Meta-Llama-3-70B](https://huggingface.co/meta-llama/Meta-Llama-3-70B)
>
> The same steps apply for any other gated model you use: visit its model card on Hugging Face, request access, accept the license, and wait for approval.

**LoRA fine-tuning example:**

Execute a basic fine-tuning example to validate the complete setup. This demonstrates parameter-efficient fine-tuning using a small model suitable for testing.
For the examples below, we are using YAML for configuration, and parameter overrides are passed as command line arguments.

```bash
## Run basic LLM fine-tuning example
cd /opt/Automodel
python3 examples/llm_finetune/finetune.py \
-c examples/llm_finetune/llama3_2/llama3_2_1b_squad_peft.yaml \
--model.pretrained_model_name_or_path meta-llama/Llama-3.1-8B \
--packed_sequence.packed_sequence_size 1024 \
--step_scheduler.max_steps 20
```

These overrides ensure the Llama-3.1-8B LoRA run behaves as expected:
- `--model.pretrained_model_name_or_path`: selects the Llama-3.1-8B model to fine-tune from the Hugging Face model hub (weights fetched via your Hugging Face token).
- `--packed_sequence.packed_sequence_size`: sets the packed sequence size to 1024 to enable packed sequence training.
- `--step_scheduler.max_steps`: sets the maximum number of training steps. We set it to 20 for demonstration purposes, please adjust this based on your needs.

> [!NOTE]
> The recipe YAML `llama3_2_1b_squad_peft.yaml` defines training hyperparameters (LoRA rank, learning rate, etc.) that are reusable across Llama model sizes. The `--model.pretrained_model_name_or_path` override determines which model weights are actually loaded.

**QLoRA fine-tuning example:**

We can use QLoRA to fine-tune large models in a memory-efficient manner.

```bash
cd /opt/Automodel
python3 examples/llm_finetune/finetune.py \
-c examples/llm_finetune/llama3_1/llama3_1_8b_squad_qlora.yaml \
--model.pretrained_model_name_or_path meta-llama/Meta-Llama-3-70B \
--loss_fn._target_ nemo_automodel.components.loss.te_parallel_ce.TEParallelCrossEntropy \
--step_scheduler.local_batch_size 1 \
--packed_sequence.packed_sequence_size 1024 \
--step_scheduler.max_steps 20
```

These overrides ensure the 70B QLoRA run behaves as expected:
- `--model.pretrained_model_name_or_path`: selects the 70B base model to fine-tune (weights fetched via your Hugging Face token).
- `--loss_fn._target_`: uses the TransformerEngine-parallel cross-entropy loss variant compatible with tensor-parallel training for large LLMs.
- `--step_scheduler.local_batch_size`: sets the per-GPU micro-batch size to 1 to fit 70B in memory; overall effective batch size is still driven by gradient accumulation and data/tensor parallel settings from the recipe.
- `--step_scheduler.max_steps`: sets the maximum number of training steps. We set it to 20 for demonstration purposes, please adjust this based on your needs.
- `--packed_sequence.packed_sequence_size`: sets the packed sequence size to 1024 to enable packed sequence training.

**Full Fine-tuning example:**

Run the following command to perform full (SFT) fine-tuning:

```bash
cd /opt/Automodel
python3 examples/llm_finetune/finetune.py \
-c examples/llm_finetune/qwen/qwen3_8b_squad_spark.yaml \
--model.pretrained_model_name_or_path Qwen/Qwen3-8B \
--step_scheduler.local_batch_size 1 \
--step_scheduler.max_steps 20 \
--packed_sequence.packed_sequence_size 1024
```

These overrides ensure the Qwen3-8B SFT run behaves as expected:
- `--model.pretrained_model_name_or_path`: selects the Qwen/Qwen3-8B model to fine-tune from the Hugging Face model hub (weights fetched via your Hugging Face token). Adjust this if you want to fine-tune a different model.
- `--step_scheduler.max_steps`: sets the maximum number of training steps. We set it to 20 for demonstration purposes, please adjust this based on your needs.
- `--step_scheduler.local_batch_size`: sets the per-GPU micro-batch size to 1 to fit in memory; overall effective batch size is still driven by gradient accumulation and data/tensor parallel settings from the recipe.
- `--packed_sequence.packed_sequence_size`: sets the packed sequence size to 1024 to enable packed sequence training.


## Step 7. Validate successful training completion

Validate the fine-tuned model by inspecting artifacts contained in the checkpoint directory.

```bash
## Inspect logs and checkpoint output.
## The LATEST is a symlink pointing to the latest checkpoint.
## The checkpoint is the one that was saved during training.
## below is an example of the expected output (username and domain-users are placeholders).
ls -lah checkpoints/LATEST/

## $ ls -lah checkpoints/LATEST/
## total 32K
## drwxr-xr-x 6 username domain-users 4.0K Oct 16 22:33 .
## drwxr-xr-x 4 username domain-users 4.0K Oct 16 22:33 ..
## -rw-r--r-- 1 username domain-users 1.6K Oct 16 22:33 config.yaml
## drwxr-xr-x 2 username domain-users 4.0K Oct 16 22:33 dataloader
## drwxr-xr-x 2 username domain-users 4.0K Oct 16 22:33 model
## drwxr-xr-x 2 username domain-users 4.0K Oct 16 22:33 optim
## drwxr-xr-x 2 username domain-users 4.0K Oct 16 22:33 rng
## -rw-r--r-- 1 username domain-users 1.3K Oct 16 22:33 step_scheduler.pt
```

## Step 8. Cleanup (Optional)

The container was launched with the `--rm` flag, so it is automatically removed when you exit. To reclaim disk space used by the Docker image, run:

> [!WARNING]
> This will remove the NeMo AutoModel image. You will need to pull it again if you want to use it later.

```bash
docker rmi nvcr.io/nvidia/nemo-automodel:26.02
```
## Step 9. Optional: Publish your fine-tuned model checkpoint on Hugging Face Hub

Publish your fine-tuned model checkpoint on Hugging Face Hub.
> [!NOTE]
> This is an optional step and is not required for using the fine-tuned model.
> It is useful if you want to share your fine-tuned model with others or use it in other projects.
> You can also use the fine-tuned model in other projects by cloning the repository and using the checkpoint.
> To use the fine-tuned model in other projects, you need to have the Hugging Face CLI installed.
> You can install the Hugging Face CLI by running `pip install huggingface_hub`.
> For more information, please refer to the [Hugging Face CLI documentation](https://huggingface.co/docs/huggingface_hub/en/guides/cli).

> [!TIP]
> You can use the `hf` command to upload the fine-tuned model checkpoint to Hugging Face Hub.
> For more information, please refer to the [Hugging Face CLI documentation](https://huggingface.co/docs/huggingface_hub/en/guides/cli).

```bash
## Publish the fine-tuned model checkpoint to Hugging Face Hub
## will be published under the namespace <your_huggingface_username>/my-cool-model, adjust name as needed.
hf upload my-cool-model checkpoints/LATEST/model
```

> [!TIP]
> The above command can fail if you don't have write permissions to the Hugging Face Hub, with the HF_TOKEN you used.
> Sample error message:
> ```bash
> user@host:/opt/Automodel$ hf upload my-cool-model checkpoints/LATEST/model
> Traceback (most recent call last):
>   File "/home/user/.local/lib/python3.10/site-packages/huggingface_hub/utils/_http.py", line 409, in hf_raise_for_status
>     response.raise_for_status()
>   File "/home/user/.local/lib/python3.10/site-packages/requests/models.py", line 1024, in raise_for_status
>     raise HTTPError(http_error_msg, response=self)
> requests.exceptions.HTTPError: 403 Client Error: Forbidden for url: https://huggingface.co/api/repos/create
> ```
> To fix this, you need to create an access token with *write* permissions, please see the Hugging Face guide [here](https://huggingface.co/docs/hub/en/security-tokens) for instructions.

## Step 10. Next steps

Begin using NeMo AutoModel for your specific fine-tuning tasks. Start with provided recipes and customize based on your model requirements and dataset.

```bash
## Copy a recipe for customization
cp examples/llm_finetune/finetune.py my_custom_training.py

## Edit configuration for your specific model and data, then run:
python3 my_custom_training.py
```

Explore the [NeMo AutoModel GitHub repository](https://github.com/NVIDIA-NeMo/Automodel) for more recipes, documentation, and community examples. Consider setting up custom datasets, experimenting with different model architectures, and scaling to multi-node distributed training for larger models.

## Troubleshooting

| Symptom | Cause | Fix |
|---------|--------|-----|
| `nvcc: command not found` | CUDA toolkit not in PATH | Add CUDA toolkit to PATH: `export PATH=/usr/local/cuda/bin:$PATH` |
| `pip install uv` permission denied | System-level pip restrictions | Use `pip3 install --user uv` and update PATH |
| GPU not detected in training | CUDA driver/runtime mismatch | Verify driver compatibility: `nvidia-smi` and reinstall CUDA if needed |
| Out of memory during training | Model too large for available GPU memory | Reduce batch size, enable gradient checkpointing, or use model parallelism |
| ARM64 package compatibility issues | Package not available for ARM architecture | Use source installation or build from source with ARM64 flags |
| Cannot access gated repo for URL | Certain HuggingFace models have restricted access | Regenerate your [HuggingFace token](https://huggingface.co/docs/hub/en/security-tokens); and request access to the [gated model](https://huggingface.co/docs/hub/en/models-gated#customize-requested-information) on your web browser |

> [!NOTE]
> DGX Spark uses a Unified Memory Architecture (UMA), which enables dynamic memory sharing between the GPU and CPU.
> With many applications still updating to take advantage of UMA, you may encounter memory issues even when within
> the memory capacity of DGX Spark. If that happens, manually flush the buffer cache with:
```bash
sudo sh -c 'sync; echo 3 > /proc/sys/vm/drop_caches'
```