How to run the turbo version on a 24G graphics card?

I have tested both the SGLang and Diffusers versions on a 24G RTX 4090 graphics card. While the model can be successfully loaded, an out-of-memory (OOM) error occurs during the inference process.

The issue mentions the SGLang and Diffusers versions, a 24G RTX 4090, and an OOM during inference, but names no files or tests. Start by reproducing the turbo-version inference setup on that hardware and comparing the memory behavior of both versions; done means documenting a confirmed way to run it or the applicable memory limitation.

Written by the indexing model from the issue text.