🎬 Wan2.2-NVFP4-Sparse

An extremely efficient Wan 2.2 14B variant: NVFP4 Quantization-Aware Step Distillation with Sparse Attention for Blackwell Architecture

📋 Table of Contents

• ✨ Features
• 🚀 Quick Start
• 🎬 Generation Results
• ⚡ Performance Comparison
• ⚠️ Notes
• 🤝 Community

✨ Features

• ⚡ 4-Step Inference: Two high-noise expert steps followed by two low-noise expert steps, enabling extremely fast Wan2.2 MoE generation on a single Blackwell GPU.
• 🎯 NVFP4 Quantization: Quantization-aware step distillation reduces memory traffic and compute cost while targeting Blackwell architecture.
• 🧩 Sparse Attention: Accelerates the costly O(n²) self-attention workload with sparse attention, reducing end-to-end latency for high-resolution video generation.
• 🔧 LightX2V Integration: Recommended runtime stack for stable deployment and best performance.
• 🚀 High-Quality Generation: Preserves the visual quality of Wan2.2-T2V/I2V-14B while dramatically improving inference speed.

🚀 Quick Start

We strongly recommend using the official LightX2V Docker image for the cleanest environment and best reproducibility.

Option A: Docker Recommended

# 1. Pull LightX2V Docker image
docker pull lightx2v/lightx2v:26052801-cu130-5090

# 2. Run single-GPU inference
# Text-to-video
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme.sh
# Image-to-video
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme.sh

# 3. Run multi-GPU sequence-parallel inference
# Text-to-video
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme_sp_parallel.sh
# Image-to-video
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme_sp_parallel.sh

Option B: Manual Installation

If Docker is not available, install the environment manually:

# 1. Install LightX2V
git clone https://github.com/ModelTC/LightX2V.git
cd LightX2V
uv pip install -v .

# 2. Install NVFP4 Kernel
pip install scikit_build_core uv
git clone https://github.com/NVIDIA/cutlass.git
cd lightx2v_kernel

MAX_JOBS=$(nproc) CMAKE_BUILD_PARALLEL_LEVEL=$(nproc) \
uv build --wheel \
  -Cbuild-dir=build . \
  -Ccmake.define.CUTLASS_PATH=/path/to/cutlass \
  --verbose --color=always --no-build-isolation

pip install dist/*whl --force-reinstall --no-deps

# 3. Run single-GPU inference
# Text-to-video
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme.sh
# Image-to-video
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme.sh

# 4. Run multi-GPU sequence-parallel inference
# Text-to-video
bash scripts/wan22/extreme/run_wan22_moe_t2v_extreme_sp_parallel.sh
# Image-to-video
bash scripts/wan22/extreme/run_wan22_moe_i2v_extreme_sp_parallel.sh

Single-GPU Scripts:

• run_wan22_moe_t2v_extreme.sh
• run_wan22_moe_i2v_extreme.sh

Multi-GPU Scripts:

• run_wan22_moe_t2v_extreme_sp_parallel.sh
• run_wan22_moe_i2v_extreme_sp_parallel.sh

🎬 Generation Results

"Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage"

Resolution	Wan2.2-T2V-14B	Wan2.2-NVFP4-Sparse
480p
720p

⚡ Performance Comparison

Test Environment: RTX 5090 Single GPU | LightX2V Framework | End-to-End Latency

Method	Task	GPU Number	Resolution	NFE	E2E Latency	Speedup
Wan2.2-T2V-14B	T2V	1	480p	40	734.0s	1.0x
Wan2.2-NVFP4-Sparse	T2V	1	480p	4	9.1s	80.7x
Wan2.2-T2V-14B	T2V	1	720p	40	2668.0s	1.0x
Wan2.2-NVFP4-Sparse	T2V	1	720p	4	22.5s	118.7x
Wan2.2-I2V-14B	I2V	1	480p	40	787.0s	1.0x
Wan2.2-NVFP4-Sparse	I2V	1	480p	4	10.7s	73.9x
Wan2.2-I2V-14B	I2V	1	720p	40	2685.0s	1.0x
Wan2.2-NVFP4-Sparse	I2V	1	720p	4	26.7s	100.5x

⚠️ Notes

System Requirements

• Required Hardware: NVIDIA RTX 50-series GPUs or other Blackwell architecture GPUs.
• Recommended Runtime: lightx2v/lightx2v:26052801-cu130-5090.

Dependencies

• Prepare Wan2.2 T5 / VAE components following the standard LightX2V Wan2.2 model structure.
• For I2V, also prepare the required image encoder components and input image according to the LightX2V Wan2.2 I2V script.
• Use Blackwell + NVFP4 kernels for optimal speed and memory efficiency.

Performance Tips

• Use the provided extreme inference script for the 4-step high-noise / low-noise expert schedule.
• Sparse attention is most beneficial at higher resolutions where self-attention dominates latency.
• Enable CPU offload only when GPU memory is limited, since offload can reduce throughput.

🤝 Community

• 🐛 Issues: GitHub Issues
• 🤗 Models: HuggingFace Hub
• 📖 Documentation: LightX2V Docs

---

If you find this project helpful, please give us a ⭐ on GitHub

For questions or issues, please open an issue on LightX2V or contact lvchengtao0319@gmail.com.