Blog

Notes I write to understand things properly — mostly LLM inference, efficiency, and compression.

• 2026.08.20 · draft

  Reusing the KV cache

  Why the keys and values a decode step reads never change.

• 2026.08.20 · draft

  How GPUs work, with the B200

  The hardware half of the story, and a prerequisite for the post below: the register / L1 / L2 / HBM hierarchy with the capacities and bandwidths that matter, why prefill saturates the tensor cores while decode starves them, and the ratio of peak compute to bandwidth that sets the ceiling.

• 2026.08.19 · draft

  How LLMs work, with Qwen3-8B

  A walkthrough of what actually happens when a decoder-only LLM answers a prompt — the decoder block, attention, multi-head and grouped-query attention, prefill vs. decode, the KV cache, and arithmetic intensity — with Qwen3-8B as the running example throughout.