Part 1 of 2 in Isolation notes

Reading inference internals

Some posts are just lists. This one is a reading list. It has headings so that the table of contents does something useful.

Serving stacks#

What the request path looks like between the socket and the first token. Batching, scheduling, KV-cache placement.

Kernels#

Where the numbers actually get multiplied, and what a kernel can and cannot see about the request it is serving.

Provenance#

Attaching an identity to a token stream and keeping it attached across tool calls.This is the part most people skip and the part that decides whether the rest of the pipeline is auditable.