Charles Dahab

Inside Tokenization: BPE, SentencePiece, Unigram and the Perils of Whitespace

Rotary Position Embeddings (RoPE): Mechanics, Scaling, and YaRN/NTK Tricks

KV Cache Engineering: Prefill, Decode, and Paged Key/Value Memory

Speculative Decoding: Draft Models, Assisted Generation, and Acceptance Kernels

Cutting latency by verifying multiple candidate tokens per step

Mixture‑of‑Experts (MoE): Routing, Capacity, and Load Balancing at Inference

Sparse activation for dense performance, without the tail latencies