Low-resource Wan deployment

On this page

This guide covers quantization, CPU/disk offloading, and VAE options for Wan2.1. Complete environment setup, then select a config compatible with your model and GPU backends. VRAM and host RAM requirements depend on weights, dimensions, frame count, and kernels; 8GB VRAM and 16/32GB RAM do not guarantee that every config will run.

Choose a configuration

Model and task	Configuration	Template defaults	Entry point
14B I2V 480P	480P disk offload	40 steps, CFG on, [480, 832]	CLI example below
14B I2V 720P	720P disk offload	40 steps, CFG off, [1280, 720]	720P launcher
1.3B T2V 480P	1.3B config	4 steps, CFG on, [480, 832]	Use --task t2v, change the model and config, and omit --image_path

Dimensions use [height, width]; the 720P template is portrait. Match sampling steps and CFG to the weights: the 4-step template needs matching distilled weights; original weights need their corresponding sampling config. The 720P launcher also fixes FP16 and --size 1280 720. Update the script when changing models or dimensions; changing only the JSON does not override these settings.

Run 480P I2V

1. Prepare a complete Wan2.1-I2V-14B-480P model directory and the quantized DiT, T5, CLIP, and Tiny VAE weights required by the config.
2. Edit the 480P JSON above, replacing dit_quantized_ckpt, t5_quantized_ckpt, clip_quantized_ckpt, and tae_path with real paths. lazy_load=true requires a directory of DiT weights saved by block; see --save_by_block in the conversion guide.
3. Install the configured backends, including sage_attn2 and fp8-vllm. If your GPU does not support them, choose matching backends and weights using the attention and quantization guides.

From the repository root, replace the model path and GPU ID, then run:

lightx2v_path="$PWD"
model_path="/path/to/Wan2.1-I2V-14B-480P"
export CUDA_VISIBLE_DEVICES=0
source scripts/base/base.sh

python -m lightx2v.infer \
  --model_cls wan2.1 \
  --task i2v \
  --model_path "$model_path" \
  --config_json configs/offload/disk/wan_i2v_phase_lazy_load_480p.json \
  --prompt "A white cat relaxing on a sunny beach." \
  --image_path assets/inputs/imgs/img_0.jpg \
  --save_result_path save_results/low_resource.mp4

Dimensions and frame count come from the JSON. View save_results/low_resource.mp4 after inference. Check peak VRAM and host RAM before increasing dimensions or frame count.

Optional adjustments

Edit these fields in the complete JSON. Follow the linked guides for detailed quantization and offload settings.

Goal	Settings and requirements
Reduce weight memory	Enable dit_quantized, t5_quantized, or clip_quantized with the matching scheme and weight path; a scheme alone does not enable quantization
Reduce VRAM	Set cpu_offload=true and t5_cpu_offload=true; host RAM must hold offloaded weights and buffers
Reduce DiT host RAM	Set lazy_load=true, use weights saved by block on an SSD, and tune num_disk_workers for disk throughput
Finer DiT offloading	Set offload_granularity="phase" to offload self-attention, cross-attention, and FFN phases within each block
Reduce intermediate memory	Enable clean_cuda_cache or rope_chunk and adjust rope_chunk_size as needed
Reduce VAE memory	Set use_tiling_vae=true; alternatively enable use_tae with tae_path, accounting for Tiny VAE's quality tradeoff
Reduce sampling steps	Use matching distilled weights and configs, rather than only reducing an original model's infer_steps

Wan2.1 Tiny VAE weights are available as taew2_1.pth. Allow disk space for base weights, quantized/distilled weights, and output files.

Further reading

• Offloading: CPU and disk offload settings.
• Quantization: weight formats, schemes, and backend dependencies.
• Gradio-Lightx2v: standalone graphical interface and deployment instructions.