One doc tagged with "single-gpu"

A memory-first method for deciding what local model configuration may fit and remain useful on one 8–24 GB accelerator.