EraserDiT: Fast Video Inpainting with Diffusion Transformer Model

git clone https://github.com/MGTVAI/EraserDiT.git
cd EraserDiT
sudo apt-get install -y git curl build-essential ffmpeg libgl1 libglib2.0-0
nvcc --version
uv venv --python 3.10
uv pip install --python .venv/bin/python --index-strategy unsafe-best-match \
  --extra-index-url https://download.pytorch.org/whl/cu126 \
  torch==2.6.0+cu126 pip setuptools wheel packaging ninja psutil
MAX_JOBS=4 uv pip install --python .venv/bin/python --index-strategy unsafe-best-match \
  --no-build-isolation-package flash-attn -r requirements.txt
uv pip check --python .venv/bin/python
HF_HUB_OFFLINE=0 uv run --no-project hf download jieeliu/EraserDiT \
  --revision 904fb412da76235085dbbccaefdbde4979fa3d29 \
  --local-dir data/model --exclude ".DS_Store"

mkdir -p outputs
export HF_HUB_OFFLINE=1 OMP_NUM_THREADS=8 MKL_NUM_THREADS=8
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
export MGERASE_NCCL_PACKING=direct MGERASE_DIT_BOUNDARY_TRANSPORT=cpu
export MGERASE_POSTPROCESS_CHUNKED_FP32=1 MGERASE_VAE_INPLACE_ACTIVATIONS=1
export MGERASE_VAE_OFFLOAD_MODE=cached MGERASE_FFMPEG_THREADS=auto
export MGERASE_ULYSSES_HEAD_CHUNKS=1 MGERASE_ULYSSES_OUTPUT_OVERLAP=0

RUNTIME=(
  --model-path data/model
  --no-dit-layerwise-offload --text-encoder-cpu-offload --vae-cpu-offload
  --vae-low-memory --cuda-memory-limit-gib 22
  --runtime-mode windowed_streaming --streaming-cache-dtype uint8
  --infer-len 121 --overlap 9 --no-compact-tail-padding
  --attention-backend sdpa
  --operator-fusion-backend triton --operator-fusion-ops qk_rmsnorm_rope,rmsnorm_adaln
  --num-inference-steps 50 --strength .8 --guidance-scale 3 --seed 42
  --transformer-cache-mode off --no-cache-text-projections
  --warmup --warmup-steps 1
)
VIDEO=(
  --video-input data/113000356.mp4 --mask-input data/113000356_mask.mp4
  --prompt "There is a bridge over the lake."
)

CUDA_VISIBLE_DEVICES=0 uv run --no-project python -m entrypoints.cli.erase_eraserdit \
  "${RUNTIME[@]}" "${VIDEO[@]}" --dit-cpu-offload --output-path outputs/l40s-1.mp4

CUDA_VISIBLE_DEVICES=0,1 uv run --no-project python -m entrypoints.cli.erase_eraserdit \
  "${RUNTIME[@]}" "${VIDEO[@]}" --no-dit-cpu-offload \
  --dit-parallel-backend nccl --cfg-degree 2 --output-path outputs/l40s-2.mp4

CUDA_VISIBLE_DEVICES=0,1,2,3 MGERASE_ULYSSES_HEAD_CHUNKS=4 MGERASE_ULYSSES_OUTPUT_OVERLAP=1 \
uv run --no-project python -m entrypoints.cli.erase_eraserdit \
  "${RUNTIME[@]}" "${VIDEO[@]}" --no-dit-cpu-offload \
  --dit-parallel-backend nccl --cfg-degree 2 --sp-degree 2 --sp-linear-mode reference \
  --output-path outputs/l40s-4.mp4

CUDA_VISIBLE_DEVICES=0 uv run --no-project python -m entrypoints.server.serve \
  --model-path data/model --task-root outputs/service --host 127.0.0.1 --port 30000 \
  --no-dit-layerwise-offload --dit-cpu-offload --text-encoder-cpu-offload --vae-cpu-offload \
  --vae-low-memory --cuda-memory-limit-gib 22 --runtime-mode windowed_streaming \
  --attention-backend sdpa --operator-fusion-backend triton \
  --operator-fusion-ops qk_rmsnorm_rope,rmsnorm_adaln --warmup --warmup-steps 2