berryber09/Wan2.2-I2V-14B-INT8-W8A8

Wan 2.2 I2V 14B — INT8 W8A8 Quantized

INT8 tensorwise quantized versions of Wan2.2-I2V-A14B for fast inference on NVIDIA GPUs with INT8 tensor cores.

Models

• High noise expert (early steps): wan2.2_i2v_high_noise_14B_fp16_learned_int8mixed_tensorwise.safetensors (~14 GB)
• Low noise expert (late steps): wan2.2_i2v_low_noise_14B_fp16_learned_int8mixed_tensorwise.safetensors (~14 GB)

Quantization Details

• Method: W8A8 (8-bit weights, 8-bit activations) with learned tensorwise scales
• Format: Mixed precision — attention/norm layers kept in FP16, linear layers quantized to INT8
• Quality: Near-lossless compared to FP16 original
• Speed: ~2x faster than FP16/BF16 on RTX 3090 (Ampere) with torch.compile

Requirements

• ComfyUI
• ComfyUI-Flux2-INT8 custom node (provides the INT8 W8A8 loader)
• NVIDIA GPU with INT8 tensor cores (RTX 20-series or newer)
• PyTorch with CUDA support
• Triton (for torch.compile acceleration)

Usage in ComfyUI

1. Place both safetensors files in ComfyUI/models/diffusion_models/
2. Use the "Load Diffusion Model INT8 (W8A8)" node to load
3. Wan 2.2 I2V workflows use both experts — high noise for early steps, low noise for late steps

Performance (RTX 3090, 24 GB)

• Model weights: ~8.9 GB on GPU (partially offloaded with --normalvram)
• Peak VRAM during sampling: ~23 GB (normal for 14B model + activations)
• Compatible with: LoRA (bypass mode), sage attention, torch.compile

INT8 is the fastest practical quantization for Ampere GPUs. FP8 requires Ada/Hopper.

Credits

• Wan-AI for the original Wan 2.2 models
• dxqb/OneTrainer for the INT8 quantization code
• bertbobson/ComfyUI-Flux2-INT8 for the ComfyUI integration