I am reading these papers:
✅ LLaMA: Open and Efficient Foundation Language Models
✅ Llama 2: Open Foundation and Fine-Tuned Chat Models
☑️ OPT: Open Pre-trained Transformer Language Models
✅ Attention Is All You Need
✅ Root Mean Square Layer Normalization
✅ GLU Variants Improve Transformer
✅ RoFormer: Enhanced Transformer with Rotary Position Embedding
✅ Self-Attention with Relative Position Representations
☑️ BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
☑️ To Fold or Not to Fold: a Necessary and Sufficient Condition on Batch-Normalization Layers Folding
✅ Fast Transformer Decoding: One Write-Head is All You Need
✅ GQA: Training Generalized Multi-Query Transformer Models from Multi-Head Checkpoints
☑️ PaLM: Scaling Language Modeling with Pathways
✅ Understand the concept of dot product of two matrices.
✅ Understand the concept of autoregressive language models.
✅ Understand the concept of attention computation.
✅ Understand the workings of Byte-Pair Encoding (BPE) algorithm and tokenizer.
✅ Read and implement the workings of the SentencePiece library and tokenizer.
✅ Understand the concept of tokenization, input ids and embedding vectors.
✅ Understand & implement the concept of positional encoding.
✅ Understand the concept of single head self-attention.
✅ Understand the concept of scaled dot-product attention.
✅ Understand & implement the concept of multi-head attention.
✅ Understand & implement the concept of layer normalization.
✅ Understand the concept of masked multi-head attention & softmax layer.
✅ Understand and implement the concept of RMSNorm and difference with LayerNorm.
✅ Understand the concept of internal covariate shift.
✅ Understand the concept and implementation of feed-forward network with ReLU activation.
✅ Understand the concept and implementation of feed-forward network with SwiGLU activation.
✅ Understand the concept of absolute positional encoding.
✅ Understand the concept of relative positional encoding.
✅ Understand and implement the rotary positional embedding.
✅ Understand and implement the transformer architecture.
✅ Understand and implement the original Llama (1) architecture.
✅ Understand the concept of multi-query attention with single KV projection.
✅ Understand and implement grouped query attention from scratch.
✅ Understand and implement the concept of KV cache.
✅ Understand and implement the concept of Llama2 architecture.
✅ Test the Llama2 implementation using the checkpoints from Meta.
✅ Download the checkpoints of Llama2 and inspect the inference code and working.
☑️ Documentation of the Llama2 implementation and repo.
✅ Work on implementation of enabling and disabling the KV cache.
✅ Add the attention mask when disabling the KV cache in Llama2.
✅ LLAMA: OPEN AND EFFICIENT LLM NOTES
✅ UNDERSTANDING KV CACHE
✅ GROUPED QUERY ATTENTION (GQA)
🌐 pytorch-llama - PyTorch implementation of LLaMA by Umar Jamil.
🌐 pytorch-transformer - PyTorch implementation of Transformer by Umar Jamil.
🌐 llama - Facebook's LLaMA implementation.
🌐 tensor2tensor - Google's transformer implementation.
🌐 rmsnorm - RMSNorm implementation.
🌐 roformer - Rotary Tranformer implementation.
🌐 xformers - Facebook's implementation.
✅ Understanding SentencePiece ([Under][Standing][_Sentence][Piece])
✅ SwiGLU: GLU Variants Improve Transformer (2020)