- Add enable_sequential_cpu_offload=true to DownloadAndLoadCogVideoModel - Reduces VRAM from ~20GB to ~12GB at cost of slower inference - Widget values: [model, precision, quantization, cpu_offload] = ['THUDM/CogVideoX-5b-I2V', 'bf16', 'disabled', true] - Necessary for 24GB GPU with other services running