notes

Paper notes and implementation writeups.

These notes track my learning process in public: one mechanism at a time, with enough implementation detail, visuals, and reproduction notes to make the code path inspectable.

  1. Why KV cache papers matter

    The opening map of the memory problem behind modern inference serving.

  2. Reproducing MiniCache in PyTorch

    The first implementation report, centered on arXiv:2405.14366.

  3. Adding a KV-cache object

    The second implementation report: applying the pairwise path to stored keys and values.

categories
next note

Next up: a decode benchmark note that makes the latency and memory tradeoff visible.