Sparse mixture of experts language model from scratch inspired by (and largely based on) Andrej Karpathy's makemore (https://github.com/karpathy/makemore) :)
HuggingFace Community Blog that walks through this: https://huggingface.co/blog/AviSoori1x/makemoe-from-scratch
This is an implementation of a sparse mixture of experts language model from scratch. This is inspired by and largely based on Andrej Karpathy's project 'makemore' and borrows the re-usable components from that implementation. Just like makemore, makeMoE is also an autoregressive character-level language model but uses the aforementioned sparse mixture of experts architecture.
Just like makemore, pytorch is the only requirement (so I hope the from scratch claim is justified).
Significant Changes from the makemore architecture
- Sparse mixture of experts instead of the solitary feed forward neural net.
- Top-k gating and noisy top-k gating implementations.
- initialization - Kaiming He initialization used here but the point of this notebook is to be hackable so you can swap in Xavier Glorot etc. and take it for a spin.
Unchanged from makemore
- The dataset, preprocessing (tokenization), and the language modeling task Andrej chose originally - generate Shakespeare-like text
- Casusal self attention implementation
- Training loop
- Inference logic
Publications heavily referenced for this implementation:
- Outrageosly Large Neural Networks: The Sparsely-Gated Mixture-Of-Experts layer: https://arxiv.org/pdf/1701.06538.pdf
- Mixtral of experts: https://arxiv.org/pdf/2401.04088.pdf
makMoE_from_Scratch.ipynb walks through the intuition for the entire model architecture and how everything comes together. I recommend starting here.
makeMoE_Concise.ipynb is the consolidated hackable implementation that I encourage you to hack, understand, improve and make your own
The code was entirely developed on Databricks using a single A100 for compute. If you're running this on Databricks, you can scale this on an arbitrarily large GPU cluster with no issues, on the cloud provider of your choice.
I chose to use MLFlow (which comes pre-installed in Databricks. It's fully open source and you can pip install easily elsewhere) as I find it helpful to track and log all the metrics necessary. This is entirely optional.
Please note that the implementation emphasizes readability and hackability vs. performance, so there are many ways in which you could improve this. Please try and let me know!
Hope you find this useful. Happy hacking!!