| # Wan2.2-I2V-A14B HiFloat4 PTQ |
|
|
| https://github.com/Reopen-AI/Wan2.2-I2V-14B-HiF4 |
|
|
| This repository contains the final optimized artifacts for the ICME 2026 |
| Low-Bit-width Large-Model Quantization Challenge, Track 1. |
|
|
| ## Base Model |
|
|
| - Base model: Wan2.2-I2V-A14B BF16 |
| - Task: image-to-video generation |
| - Quantization: HiFloat4 W4A4 quant-dequant simulation |
|
|
| ## Quantization Policy |
|
|
| - `low_noise_model`: quantized |
| - `high_noise_model`: quantized |
| - First and last Transformer blocks: BF16 |
| - Other `nn.Linear` layers: HiFloat4 W4A4 QDQ |
| - Activations: HiFloat4 QDQ enabled |
| - Exported weights: BF16 tensors after HiFloat4 QDQ, with SmoothQuant and clip |
| metadata |
|
|
| ## Files |
|
|
| ```text |
| i2v_hif4_quant_state.pt |
| i2v_hif4_quantized_weights_low_noise_model.pt |
| i2v_hif4_quantized_weights_high_noise_model.pt |
| ``` |
|
|
| `i2v_hif4_quant_state.pt` stores calibration metadata. The two exported-weight |
| files store the low-noise and high-noise expert weights separately. |
|
|
| ## Usage |
|
|
| Clone the code repository and place or download these files under `outputs/`. |
| Then run: |
|
|
| ```bash |
| LOW_QUANT_WEIGHTS=outputs/i2v_hif4_quantized_weights_low_noise_model.pt \ |
| HIGH_QUANT_WEIGHTS=outputs/i2v_hif4_quantized_weights_high_noise_model.pt \ |
| OUT_DIR=/path/to/generated_videos \ |
| bash opens2v_generate.sh |
| ``` |
|
|
| For full reproduction instructions, see the code repository README. |
|
|
