CogVideo demo

This is an unofficial demo app for CogVideo.

You can try web demo: (This version currently supports only the first stage.)

sample_00.mp4

sample_01.mp4

It takes about 7 minutes to load models on startup and about 11 minutes to generate one video.

Prerequisite

An A100 instance is required to run CogVideo.

Installation

Change default-runtime of docker

First, put "default-runtime": "nvidia" in /etc/docker/daemon.json. See: NVIDIA/nvidia-docker#1033 (comment)

{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

Then, restart docker.

sudo systemctl restart docker

Clone this repo

git clone --recursive https://github.com/hysts/CogVideo_demo
cd CogVideo_demo

Build docker image

docker build . -t cogvideo

Apply patch to CogVideo repo

cd CogVideo
patch -p1 < ../patch

Download pretrained models (Optional)

The pretrained models will be downloaded automatically on the first run, but it may take quite some time. So you may want to download them in advance.

This repo assumes the pretrained models are stored in the pretrained directory as follows:

pretrained
├── cogvideo-stage1
│   ├── 27000
│   │   └── mp_rank_00_model_states.pt
│   ├── latest
│   └── model_config.json
├── cogvideo-stage2
│   ├── 38000
│   │   └── mp_rank_00_model_states.pt
│   ├── latest
│   └── model_config.json
└── cogview2-dsr
     ├── 20000
     │   └── mp_rank_00_model_states.pt
     ├── latest
     └── model_config.json

Run

You can run the app with the following command:

docker compose run --rm app

The app will start up on port 7860 by default. You can change the port using GRADIO_SERVER_PORT environment variable. Use port forwarding when running on GCP, etc.