nanoGPT_mlx

A port of Andrej Karpathy's nanoGPT in Apple's new machine learning framework, MLX.

Train OpenAI's GPT-2 models or custom GPT-style models from scratch, all on your Mac's GPU!

Still under active development, but currently the file train.py closely resembles the nanoGPT codebase.

install

Create a conda environment using the provided environment configuration file.

conda env create -f environment.yaml

Activate conda environment.

conda activate apple_mlx

Dependencies:

• mlx
• numpy
• datasets for huggingface datasets (if you want to download + preprocess OpenWebText)
• tiktoken for OpenAI's fast BPE code
• tensorboardX for optional logging
• tqdm for progress bars

quick start

To train a character-level GPT, prepare shakespeare dataset similar to nanoGPT. This will create a train.bin and val.bin in that data directory.

python data/shakespeare/prepare.py

Now, let's train a "baby GPT" model on your MAC GPU:

python train.py configs/train_gpt2_shakespeare.py

On my Macbook M3 Pro, I am observing ~0.37 iterations/second when training a ~45M parameter GPT-2 model, at batch_size=64 (i.e., local_batch_size=4 and gradient_accumulation=16).

So once the training finishes we can sample from the best model by pointing the sampling script at this directory:

python sample.py --out_dir=gpt2_small_shakespeare

openwebtext

To train a GPT-2 model on OpenWebText similar to nanoGPT, first prepare the dataset:

python data/openwebtext/prepare.py

Then, train a 124M GPT-2 model on your MAC GPU:

python train.py configs/train_gpt2_owt.py

todos

• disable weight decay on non-decay params in optimizer
• add bfloat16 training support
• integrate Eleuther Eval
• add checkpoint conversion for loading pre-trained HF models
• add saveing and loading pre-trained MLX models
• enable finetuning models from pre-trained checkpoints
• enable inference with pre-trained models

acknowledgements

Thank you Andrej Karpthy for creating the nanoGPT codebase. It's been awesome for quick prototyping!