ApacheOne/Wan2.2-Animate-2-14B-OrbitQuant-W4A4

Wan2.2 Animate-2 14B Distilled — OrbitQuant W4A4

Packed OrbitQuant W4A4 transformer for:

Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers

Quantization

This artifact uses:

• 4-bit packed nonuniform OrbitQuant weights
• online 4-bit OrbitQuant activations
• RPBH rotation
• exact-density Lloyd-Max codebooks
• BF16 per-row weight scales
• custom Triton packed W4 × A4 GEMM
• no dense target transformer weights at runtime

Transformer architecture:

• 40 layers
• hidden dimension: 5120
• FFN dimension: 13824
• 40 attention heads
• 480 packed target linear layers
• 823 non-target tensors retained
• 1303 transformer tensors total

Files

The twelve:

orbitquant-runtime-00001.safetensors through orbitquant-runtime-00012.safetensors

contain the packed transformer and non-target tensors.

orbitquant_rotations.safetensors contains the OrbitQuant rotation / codebook data.

packed_manifest.json describes every packed target, tensor location, shape, packed layout, row scale, quantization statistics, and audit.

Runtime

The runtime/ directory contains the custom OrbitQuant loader and Triton W4A4 runtime used for validation.

This is not a normal Diffusers dense transformer checkpoint and requires the included OrbitQuant runtime.

Other Wan-Animate-2 components

This repository contains the OrbitQuant transformer artifact.

Text encoder, image encoder, VAE, tokenizers, and other non-transformer components should be obtained from the upstream Wan-Animate-2 model.

Validation

Validated end-to-end on an NVIDIA L4 using CUDA 12.8.

The runtime CUDA gate verified:

• RPBH activation rotation
• online Lloyd-Max A4 packing
• packed nonuniform W4 × A4 Triton GEMM
• actual packed transformer matrices

All 480 packed target modules were exercised during end-to-end Wan-Animate-2 inference.

See:

validation/orbitquant_w4a4_cuda_gate.json

for the recorded kernel validation data.