LLM inference

Zhipu built its own inference stack for GLM

GLM built its own inference infrastructure to serve LLMs cheaply on constrained, mixed hardware. Here's what breaks in generic stacks and what to copy.