10-12-2026, 01:33 PM
Thanks, that helps. I ended up in about the same place using vLLM. The biggest change for me was simplifying the stack instead of adding one more optimization layer.
Mostly CUDA, sometimes wrong, usually measured
