Instructions to use nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForMultimodalLM

tokenizer = AutoTokenizer.from_pretrained("nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1")
model = AutoModelForMultimodalLM.from_pretrained("nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1

SGLang

How to use nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 with Docker Model Runner:
```
docker model run hf.co/nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1
```

Pruning doesn't appear to be viable.

by phil111 - opened May 26, 2025

Discussion

phil111

May 26, 2025

How did Nvidia come to the conclusion that pruning is a good idea? (https://arxiv.org/abs/2408.11796)

I tested pretty much all of your pruned models, and the weights are effectively destroyed across the board relative to the source models. Sure, if you then re-train the pruned models on a corpus focused on select domains you can boost scores in said domains, such as on the MMLU, GSM8k, and HumanEval tests, yet everything else is left far more scrambled than the source models.

For example, the source model (Llama 3.1 8b) can get the main casts of pretty much every top 100 movie and TV show correct without a single hallucinations. But this pruned, then selectively overfit model, reliably makes egregious hallucinations on all the same movies and TV shows, including the most well known movies & shows globally (e.g. Friends). Same goes for music, games, and so on.

So I fail to see how pruning and selectively overfitting models using a tiny corpus accomplishes anything beyond boosting select abilities and test scores. Why doesn't Nvidia just train their own grossly overfit test boosting LLMs from scratch? Why lobotomize existing models?

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment