From CPU-offloaded inference to a targeted W4 embedding and output-head quantization on a single RTX PRO…