ML 笔记
从 prompt 到 token 之间发生的一切，一次讲清一个机制：一张可交互的大模型推理全链路地图，从注意力、KV cache 到智能体和 GPU 算子。
What it is
Notes on how large language models actually run, written the way I wish I had learned it. Each mechanism gets its own step-through page instead of a wall of text, and every page sits on one map of the stack, so you can see where it fits.
Four tracks
- Infrastructure — from prompt to token: prefill and decode, the KV cache, batching, and what happens when one GPU is not enough.
- Agents — the loop that wraps the model: context, tools and memory.
- Blackwell — one CUDA kernel, taken from 600 microseconds to 57.
- Foundations — from a single neuron to a trained network.