gemneye/imontage repository overview

iMontage on RunPod (DockerHub image) This container is built for RunPod Pods/Serverless. It bootstraps iMontage at runtime so you always get the latest code from the fork.

Image gemneye/imontage:latest (each commit also has a gemneye/imontage: tag) What the container does at start Installs Python deps (Torch 2.7.0/cu128, torchvision 0.22.0, torchaudio 2.7.0, flash-attn 3.0.0.post1). Clones the repo into /workspace/iMontage if it is not present. pip install -e . to bring in project deps. Launches the Gradio UI (gradio_app.py) on port 7860. Required downloads (models) Before running inference you must download checkpoints per the project README into ckpts/ (e.g., ckpts/hyvideo_ckpts/..., ckpts/iMontage_ckpts/...). You can bake this into a volume or perform the download inside the container startup if desired.

Environment overrides TORCH_VERSION, TORCHVISION_VERSION, TORCHAUDIO_VERSION, TORCH_INDEX_URL – pick different torch wheels if needed. FLASH_ATTN_VERSION (flash-attn required; avoid setting INSTALL_FLASH_ATTN=0 unless you know it’s safe to skip). REPO_URL, REPO_DIR – point to a different fork/location. MODEL_BASE, DIT_WEIGHT – override checkpoint paths used by the Gradio app. Running on RunPod Pods Use gemneye/imontage:latest as the image. Expose port 7860 (container port) and map to your chosen proxy port in the template. Ensure a volume or startup command downloads the required ckpts before first inference. The default command runs the Gradio UI. To run the CLI instead, override the container command with something like: ["python3.12", "fastvideo/sample/sample_imontage.py", "--prompt", "assets/prompt.json", ...] Running locally docker run --gpus all -it -p 7860:7860 --ipc=host gemneye/imontage:latest

⁠Gradio UI will be at http://localhost:7860⁠