mahakal001 / nanogpt-jax

An implementation of nanogpt in jax from scratch based on Andrej Karpathy's tutorial series on Neural Networks.

Nano GPT-jax

An implementation of nanogpt in jax from scratch ( Other than Optax for optimization and Equinox for handling PyTrees ) based on Andrej Karpathy's Let's build GPT Lecture.

Usage

• The Shakespeare dataset is in data folder. You only need to configure hyper-parameters in nanogpt-jax/train.py as per your test settings and then run :

$ python train.py

TODOS

• Write DropOut Layers.
• LayerNorm.
• Apply weight initializers.
• Implement Adam.

References

• Andrej Karpathy's Let's build GPT Lecture.
• From PyTorch to JAX: towards neural net frameworks that purify stateful code .
• For my usecase I did not want to use Haiku or Flax. I wanted something very mimimal. And I found Equinox suitable. I got introduced to Equinox through this Repo by Phil Wang.