Layer-wise inferencing + batching: Small VRAM doesn't limit LLM throughput anymore