Explain how a GPU stores model weights and the key/value cache during language model inference. Use plain language and describe why longer context uses more memory. Give a concrete example without inventing benchmark results.
