| { |
| "quantization_method": "staged_wan_mixed_precision_nf4", |
| "model_type": "wan", |
| "description": "Wan high-noise and low-noise transformer experts are quantized in separate ZeroGPU allocations. First/last blocks and boundary modules stay in bfloat16.", |
| "stages": { |
| "high": { |
| "subfolder": "high_noise_model", |
| "layout": "original", |
| "quantization_method": "mixed_precision_nf4", |
| "high_precision_layers_count": 26, |
| "high_precision_layers": [ |
| "blocks.0.self_attn.q", |
| "blocks.0.self_attn.k", |
| "blocks.0.self_attn.v", |
| "blocks.0.self_attn.o", |
| "blocks.0.cross_attn.q", |
| "blocks.0.cross_attn.k", |
| "blocks.0.cross_attn.v", |
| "blocks.0.cross_attn.o", |
| "blocks.0.ffn.0", |
| "blocks.0.ffn.2", |
| "blocks.39.self_attn.q", |
| "blocks.39.self_attn.k", |
| "blocks.39.self_attn.v", |
| "blocks.39.self_attn.o", |
| "blocks.39.cross_attn.q", |
| "blocks.39.cross_attn.k", |
| "blocks.39.cross_attn.v", |
| "blocks.39.cross_attn.o", |
| "blocks.39.ffn.0", |
| "blocks.39.ffn.2", |
| "text_embedding.0", |
| "text_embedding.2", |
| "time_embedding.0", |
| "time_embedding.2", |
| "time_projection.1", |
| "head.head" |
| ], |
| "removed_stale_checkpoint_indexes": [ |
| "diffusion_pytorch_model.safetensors.index.json" |
| ] |
| }, |
| "low": { |
| "subfolder": "low_noise_model", |
| "layout": "original", |
| "quantization_method": "mixed_precision_nf4", |
| "high_precision_layers_count": 26, |
| "high_precision_layers": [ |
| "blocks.0.self_attn.q", |
| "blocks.0.self_attn.k", |
| "blocks.0.self_attn.v", |
| "blocks.0.self_attn.o", |
| "blocks.0.cross_attn.q", |
| "blocks.0.cross_attn.k", |
| "blocks.0.cross_attn.v", |
| "blocks.0.cross_attn.o", |
| "blocks.0.ffn.0", |
| "blocks.0.ffn.2", |
| "blocks.39.self_attn.q", |
| "blocks.39.self_attn.k", |
| "blocks.39.self_attn.v", |
| "blocks.39.self_attn.o", |
| "blocks.39.cross_attn.q", |
| "blocks.39.cross_attn.k", |
| "blocks.39.cross_attn.v", |
| "blocks.39.cross_attn.o", |
| "blocks.39.ffn.0", |
| "blocks.39.ffn.2", |
| "text_embedding.0", |
| "text_embedding.2", |
| "time_embedding.0", |
| "time_embedding.2", |
| "time_projection.1", |
| "head.head" |
| ], |
| "removed_stale_checkpoint_indexes": [ |
| "diffusion_pytorch_model.safetensors.index.json" |
| ] |
| }, |
| "text_encoder": { |
| "quantization_method": "skipped", |
| "reason": "Original Wan ships the UMT5 encoder as models_t5_umt5-xxl-enc-bf16.pth. Pre-quantizing that file would require changing the original Wan runtime loader, so it is preserved to keep the target repo usable." |
| } |
| } |
| } |
