🧪 FastVideo & DMD Distillation: Development, Debugging, and Reproduction

This guide provides streamlined instructions for working with the FastVideo distillation and DMD (CausVid) repositories, including development/debug branches, environment setup, dataset/model preparation, and execution scripts.

🔧 FastVideo Distill

🔗 Branches

Development branch:
yq/distill-dmd
Debug branch (fixed seed):
yq/distill-dmd-fixseed

🛠️ Environment Setup

Follow the official FastVideo installation instructions:
📄 Create and activate a conda environment

📁 Dataset Preparation

Ensure the dataset is available by running:

python scripts/huggingface/download_hf.py \
  --repo_id=FastVideo/mini_i2v_dataset \
  --local_dir=mini_i2v_dataset \
  --repo_type=dataset

This downloads a minimal example dataset sufficient for running the distillation script.

🚀 Distillation Script

Use the same script for both distill-dmd and distill-dmd-fixseed:

📄 v1_distill_dmd_wan.sh

🐛 Debugging Tip

To set a breakpoint in fastvideo codebase:

torch.distributed.breakpoint()

🧩 DMD Official Repo (Forked)

🔗 Branches

Reproduce branch:
yq/debug
Debug branch (with fixed seed):
yq/debug-fixseed

🛠️ Environment Setup

For yq/debug (reproduce branch):
Follow the official DMD README instructions.
For yq/debug-fixseed:
Use the same environment as FastVideo and then:

# Inside the FastVideo directory
python setup.py develop

# Then install any missing packages
pip install <missing-packages>

📦 Model & Dataset

Download the official Wan base model and dataset from Hugging Face:
📦 mixkit_latents_lmdb
➡️ Place the dataset under checkpoints/.

🚀 Running Scripts

Reproduce branch script (8-node run):
📄 dmd_8n_simple.sh
Debug branch script (fixed seed and input):
📄 debug.sh