| { |
| "strict_load": true, |
| "source_tensor_bytes_equal": true, |
| "checks": [ |
| { |
| "name": "571_source_tensor_payloads", |
| "passed": true |
| }, |
| { |
| "name": "native_visual_strict_load", |
| "passed": true |
| }, |
| { |
| "name": "complete_semantic_and_flow_layers", |
| "passed": true |
| }, |
| { |
| "name": "8x8_dual_path_fusion", |
| "passed": true |
| }, |
| { |
| "name": "8x8_native_forward_parity", |
| "passed": true, |
| "max_abs": 0.0 |
| }, |
| { |
| "name": "8x8_finite_velocity", |
| "passed": true |
| }, |
| { |
| "name": "8x12_dual_path_fusion", |
| "passed": true |
| }, |
| { |
| "name": "8x12_native_forward_parity", |
| "passed": true, |
| "max_abs": 0.0 |
| }, |
| { |
| "name": "8x12_finite_velocity", |
| "passed": true |
| }, |
| { |
| "name": "fixed_27x27_position_path", |
| "passed": true |
| }, |
| { |
| "name": "Wan21_official_LFS_SHA256", |
| "passed": true |
| }, |
| { |
| "name": "Wan21_native_strict_source_state", |
| "passed": true, |
| "tensors": 194 |
| }, |
| { |
| "name": "Wan21_1frame_encode", |
| "passed": true, |
| "shape": [ |
| 1, |
| 16, |
| 1, |
| 8, |
| 8 |
| ] |
| }, |
| { |
| "name": "Wan21_1frame_decode", |
| "passed": true, |
| "shape": [ |
| 1, |
| 3, |
| 1, |
| 64, |
| 64 |
| ] |
| }, |
| { |
| "name": "RGB_VAE_visual_boundary", |
| "passed": true |
| }, |
| { |
| "name": "Wan21_5frame_encode", |
| "passed": true, |
| "shape": [ |
| 1, |
| 16, |
| 2, |
| 4, |
| 4 |
| ] |
| }, |
| { |
| "name": "Wan21_5frame_decode", |
| "passed": true, |
| "shape": [ |
| 1, |
| 3, |
| 5, |
| 32, |
| 32 |
| ] |
| } |
| ], |
| "smoke": { |
| "passed": true, |
| "device": "cpu", |
| "dtype": "float32" |
| }, |
| "elapsed_seconds": 28.197694792004768, |
| "limitations": [ |
| "Visual forward parity uses unchanged official source methods with an echo language-backbone stub; it does not validate the excluded Qwen model, language-conditioned generation, image quality or VQA.", |
| "Visual adapter smoke coverage is single-image, small square/rectangular latent grids. Wan2.1 codec checks cover one and five frames. No full video flow head or complete sampler benchmark.", |
| "Native output token ordering and trained time-token convention are preserved. Full-model integration requires the proper multimodal sequence mask and actual final Qwen hidden states." |
| ] |
| } |
|
|
