✅ Jupyter + Hugging Face Local Model Setup (Mac M3 Optimized)

This guide sets up Python 3.11, Jupyter Notebook, Hugging Face models, and local model loading using uv and pyenv.

---

1. Install Prerequisites

Install Homebrew (if not installed)

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

Install pyenv and uv using Homebrew

brew install pyenv uv

Configure pyenv in shell

echo 'export PYENV_ROOT="$HOME/.pyenv"' >> ~/.zprofile
echo 'export PATH="$PYENV_ROOT/bin:$PATH"' >> ~/.zprofile
echo 'eval "$(pyenv init --path)"' >> ~/.zprofile
source ~/.zprofile

---

2. Install Python 3.11.9

pyenv install 3.11.9
pyenv global 3.11.9

Confirm Python

which python
# should be ~/.pyenv/shims/python
python --version
# should be Python 3.11.9

---

3. Create Project and .venv with uv

mkdir -p ~/llm/jupyter
cd ~/llm/jupyter
uv venv

---

4. Add requirements.txt

Save the following file as requirements.txt in the project folder:

torch==2.3.0
numpy==1.26.4
transformers==4.52.4
huggingface-hub==0.33.0
datasets==3.6.0  
trl==0.14.0
jinja2==3.1.2
markupsafe==2.0.1
tabulate==0.9.0
pandas==2.3.0

---

5. Install all dependencies via uv

uv pip install -r requirements.txt

✅ Verify all packages are installed

uv pip list

Look for: torch, transformers, pandas, etc.

---

6. Launch Jupyter Notebook

source .venv/bin/activate
jupyter notebook

---

7. Download Model from Hugging Face (CLI option)

First, log in (if required)

huggingface-cli login

Then download the model

huggingface-cli download Qwen/Qwen3-0.6B-Base --local-dir ./models/Qwen/Qwen3-0.6B-Base --local-dir-use-symlinks False

---

8. Load Model in Notebook

from transformers import AutoModelForCausalLM, AutoTokenizer
from pathlib import Path

def load_model_and_tokenizer(model_path, use_gpu=True):
    resolved_path = str(Path(model_path).resolve())

    model = AutoModelForCausalLM.from_pretrained(
        resolved_path,
        local_files_only=True,
        device_map="auto" if use_gpu else None
    )

    tokenizer = AutoTokenizer.from_pretrained(
        resolved_path,
        local_files_only=True
    )

    return model, tokenizer

model, tokenizer = load_model_and_tokenizer("./models/Qwen/Qwen3-0.6B-Base", use_gpu=False)

---

9. Test Inference

input_ids = tokenizer("Hello, world!", return_tensors="pt").input_ids
output = model.generate(input_ids)
print(tokenizer.decode(output[0], skip_special_tokens=True))

---

✅ You’re Ready!

• Runs fast on Mac M3
• No remote HF calls
• Jupyter is isolated in .venv
• Fully reproducible with requirements.txt