Hacker News Reader

> GPUs are extremely underutilized if you launch just 1 generation stream

why is that? b/c the thing is waiting for the hoooman and idling? or some parallelizable interleaving steps?

I have no intuition yet how this works under the hood.

Some of the inference engines can process multiple requests in parallel more efficiently than doing them sequentially. Not sure of the exact mechanism but e.g. llama.cpp's llama-server can do this (you tell it the number of slots to have when starting, then fire HTTP requests at it and it batches them together when it can).

Waiting for the hooman (or tool calls) won't help either, of course.

The mechanism is that generating tokens (the "decode" phase) in an LLM is limited by memory bandwidth for the weights, so computing multiple streams amortizes the bandwidth over streams as long as you can keep the contexts in RAM. This is most true for dense models and the always-on expert in MoE models, or when you have significantly more streams than experts for MoE models.

In contrast, prompt prefill is more easily compute-bound, so there are interesting trade-offs for latency of decode vs prefill when the LLM utilization is high.

Reply on news.ycombinator.com