MLX, PyTorch & Hugging Face Demo on MacBook Pro (M4 Pro+)

Introduction

This project demonstrates how to leverage Apple's powerful M4 Pro chip for machine learning workloads using three complementary frameworks:

• MLX: Apple's native machine learning framework optimized for Apple Silicon
• PyTorch with MPS: The popular deep learning framework with Metal Performance Shaders support
• Hugging Face: Access to state-of-the-art pre-trained models and transformers

Why This Combination?

The MacBook M4 Pro offers exceptional performance for machine learning tasks when properly configured:

🚀 Apple MLX Framework

• Native optimization for Apple Silicon architecture
• Unified memory access for efficient model training and inference
• Seamless integration with Metal Performance Shaders
• Optimized for transformer models and large language models

⚡ PyTorch with MPS Backend

• Industry-standard deep learning framework
• Metal Performance Shaders (MPS) acceleration on Apple Silicon
• Broad ecosystem support and extensive documentation
• Easy migration from CUDA-based workflows

🤗 Hugging Face Integration

• Access to thousands of pre-trained models
• Streamlined model downloading and management
• Compatible with both MLX and PyTorch backends
• Community-driven model repository

Hardware Advantages of M4 Pro

The MacBook M4 Pro provides several advantages for ML workloads:

• Unified Memory Architecture: Shared memory between CPU and GPU eliminates data transfer bottlenecks
• High Memory Bandwidth: Up to 273 GB/s memory bandwidth for large model handling
• Neural Engine: Dedicated 16-core Neural Engine for optimized ML operations
• Energy Efficiency: Superior performance-per-watt compared to traditional GPU setups

Prerequisites

• MacBook with M4 Pro chip (or other Apple Silicon)
• macOS 14.0 or later
• Python 3.9+
• UV package manager (recommended) or pip

Installation

1. Install Required Packages

# Install MLX, transfer learning utilities, and tokenization support
$ uv add mlx tf-transfer sentencepiece

# Install Hugging Face CLI tools
$ uv tool install "huggingface-hub[cli]"

# Install PyTorch with MPS support (if not already installed)
$ uv add torch torchvision torchaudio

2. Download Pre-trained Models

# Download Mistral 7B model optimized for MLX
$ uvx -w "hf-transfer" --from "huggingface-hub[cli]" huggingface-cli download \
  --local-dir-use-symlinks False \
  --local-dir mistral-7B-v0.1 \
  mlx-model/mistral-7B-v0.1

Features Demonstrated

This project showcases:

• ✅ Device Detection: Automatic detection and utilization of Apple Silicon GPU via MPS
• ✅ MLX Integration: Native Apple Silicon optimization for ML workloads
• ✅ Model Loading: Efficient loading of Hugging Face models
• ✅ Performance Comparison: Benchmarking between CPU, MPS, and MLX backends
• ✅ Memory Management: Optimal memory usage with unified memory architecture

Project Structure

torch-mps-demo/
├── main.py              # Main demonstration script
├── README.md            # This file
├── mistral-7B-v0.1/     # Downloaded model files (created after setup)
└── requirements.txt     # Python dependencies (if using pip)

Usage

Run the main demonstration:

python main.py

This will automatically detect your hardware capabilities and run appropriate tests for:

• PyTorch MPS backend functionality
• MLX framework integration
• Hugging Face model loading and inference

Performance Tips

1. Memory Management: Close unnecessary applications to maximize available unified memory
2. Thermal Management: Ensure proper ventilation for sustained performance
3. Model Sizing: Choose model sizes appropriate for your M4 Pro's memory configuration
4. Batch Processing: Optimize batch sizes for your specific hardware configuration