Wan2.2-I2V-A14B-Diffusers-MXFP8-AutoRound

An AutoRound MXFP8 checkpoint of Wan-AI/Wan2.2-I2V-A14B-Diffusers for image-to-video generation with vLLM-Omni.

Both diffusion experts (transformer and transformer_2) are quantized. The text encoder, tokenizer, VAE, and scheduler are included. Parameters outside the quantized layers keep their original precision.

Quantization

Setting	Value
Method	AutoRound, round-to-nearest (`--iters 0 --disable_opt_rtn`)
Scheme	MXFP8
Weight storage	FP8 E4M3 (`F8_E4M3` in safetensors)
Weight scale storage	`uint8` MX block scales
Weight / activation group size	32 / 32
Quantized blocks	`blocks`, in both experts
Packing format	`auto_round:llm_compressor`
AutoRound version	0.17.0

The checkpoint was produced with:

auto-round \
  --model ./Wan2.2-I2V-A14B-Diffusers \
  --scheme mxfp8 \
  --iters 0 \
  --disable_opt_rtn \
  --output_dir ./Wan2.2-I2V-A14B-Diffusers-MXFP8-AutoRound

vLLM-Omni inference

Install vLLM-Omni with offline AutoRound MXFP8 support and compatible dependencies. Activate your environment, then run this command from the vLLM-Omni source directory. Download assets/input.png from this repository, or use your own input image.

720p generation

MODEL=intel-ai/Wan2.2-I2V-A14B-Diffusers-MXFP8-AutoRound

python \
  examples/offline_inference/image_to_video/image_to_video.py \
  --model "$MODEL" \
  --image /path/to/assets/input.png \
  --prompt "A cinematic shot of the scene, gentle natural motion, smooth camera movement, high quality." \
  --height 720 \
  --width 1280 \
  --num-frames 81 \
  --num-inference-steps 50 \
  --guidance-scale 5.0 \
  --boundary-ratio 0.9 \
  --flow-shift 5.0 \
  --seed 42 \
  --fps 16 \
  --enforce-eager \
  --vae-use-tiling \
  --output mxfp8.mp4

AutoRound quantization is automatically detected from the checkpoint. No extra quantization flag is needed. You can also set MODEL to a local path.

BF16 comparison

To generate the BF16 reference, run the same command with MODEL=Wan-AI/Wan2.2-I2V-A14B-Diffusers and --output bf16.mp4. Keep the image, prompt, seed, and all sampling settings the same.

Example output

Both examples use 1280 × 720, 81 frames, 50 inference steps, and 16 FPS. The input image comes from examples/i2v_input.JPG in the base model, saved here as PNG without changing the decoded pixels.