Blog
Notes on LLM internals and the systems that serve them.
Understanding KV Cache
Why KV cache exists, how decode differs from prefill, and why LLM serving starts looking a lot like operating systems.
How Transformers work and why Attention is the key
Why RNNs hit a wall, how self-attention and multi-head attention actually work under the hood, and what happens during a full encoder-decoder forward pass.