6 docs tagged with "deep-dive"

Q·K·V, Scaled Dot-Product Attention, multi-head, causal masking, and the Transformer block — a technical look at how an LLM computes

How full fine-tuning and PEFT (LoRA/QLoRA) work, and how adapters update weights in a low-rank way

KV cache, quantization, batching, FlashAttention, GQA/MQA and more — techniques to serve LLMs fast and cheap

How RAG retrieves and incorporates sources — chunking, embeddings, vector similarity, ANN search, reranking, context injection

How the next token is chosen from logits — temperature, top-k, top-p (nucleus), repetition penalty, greedy and beam search

How an LLM splits text into tokens (BPE / WordPiece / SentencePiece), vocabulary, and why Korean uses more tokens