Memory-Efficient Backpropagation for Fine-Tuning LLMs on Resource-Constrained Mobile Devices

This is a demo implementation for Memory-Efficient Backpropagation for Fine-Tuning LLMs on Resource-Constrained Mobile Devices .

The implementation is based on MLX and includes:

1. Exporting Qwen3 model to MLX graphs for fine-tuning with QLoRA.
2. On-device Memory-Efficient Backpropagation (MeBP) runtime implementation.
3. Example iOS App for running MeBP on Qwen3.

1. Run notebook in examples/export_qwen3.ipynb to export Qwen3 model.
2. The notebook will generate model assets at output directory specified (default to data/qwen3)

1. Connect your iPhone to your MacBook
2. Open MLXMeBPExample/MLXMeBPExample.xcodeproj in Xcode
3. Build the App on your iPhone

1. Open Finder App on your MacBook and find your connected iPhone on the left
2. Click Files in Finder which shows a list of Apps installed on your iPhone
3. Find MLXMeBPExample App and drag the Qwen3 directory to the Apps' directory

1. Open MLXMeBPExample App on your iPhone
2. Click Initialize Model
3. Click Start Training