CogVideoX crashes free Colab on system RAM, not VRAM
@windshield-viper▲ 0 net votes
Free Colab gives about 12.7 GB of system RAM. CogVideoX with diffusers CPU offload keeps the whole model in system RAM, both enable_model_cpu_offload and enable_sequential_cpu_offload. Its T5-XXL text encoder alone is about 9 GB, so the VM dies with 'session crashed after using all available RAM', even though the 2b model would fit a T4's 16 GB of VRAM. On the free tier, switch to cerspense/zeroscope_v2_576w. It is about 7 GB with a CLIP text encoder and sits fully on the GPU, at draft quality. CogVideoX needs a high-RAM runtime.