This is a repo of my experiments and notes while learning about LLMs. I'm starting with a decent theoretical understanding of neural networks, and hands on experience training large models on distributed systems. I'm very comfortable with data and ML engineering.

I've read many papers and a few books in the deep learning and LLM space, but have never committed to learning things deeply or hands on. I plan to change that.

Here are all the things I'd like to do:

• Andrej Karpathy's Neural Networks: Zero to Hero guide.
• Thoroughly read The Annotated Transformers paper and run the code side-by-side.
• Explore the Tensor2Tensor repo.

• Implement FlashAttention myself (in Cuda maybe?)
• Implement FSDP myself (no idea how!?)

• Model efficiency experiments. Try out the following and benchmark performance changes:
  • Speculative decoding
  • Knowledge distillation
  • Quantization
  • Pruning
  • Sparsity low ran compression
  • etc
• Play around with LLAMA models locally

• Reread: Flash Attention 1, 2 and Paged Attention
• Flash Attention 3
• Depthwise Seperable Convolutions for NMT
• One Model To Learn Them All
• Self-Attention with Relative Position Representations
• Self-attention Does Not Need O(n2) Memory
• Oneline softmax papers