| { |
| "strict_load": true, |
| "source_tensor_bytes_equal": true, |
| "checks": [ |
| { |
| "name": "1190_original_tensor_payloads", |
| "passed": true |
| }, |
| { |
| "name": "strict_all_1190_entries", |
| "passed": true |
| }, |
| { |
| "name": "flow_scaling_factor", |
| "passed": true |
| }, |
| { |
| "name": "last_six_Qwen_layers", |
| "passed": true |
| }, |
| { |
| "name": "112px_image_features", |
| "passed": true |
| }, |
| { |
| "name": "112px_shuffle_order_exact", |
| "passed": true |
| }, |
| { |
| "name": "112px_native_latent", |
| "passed": true |
| }, |
| { |
| "name": "112px_native_decoder_exact", |
| "passed": true |
| }, |
| { |
| "name": "112px_finite_RGB", |
| "passed": true, |
| "shape": [ |
| 1, |
| 3, |
| 128, |
| 128 |
| ] |
| }, |
| { |
| "name": "448px_image_features", |
| "passed": true |
| }, |
| { |
| "name": "448px_shuffle_order_exact", |
| "passed": true |
| }, |
| { |
| "name": "448px_native_latent", |
| "passed": true |
| }, |
| { |
| "name": "448px_native_decoder_exact", |
| "passed": true |
| }, |
| { |
| "name": "448px_finite_RGB", |
| "passed": true, |
| "shape": [ |
| 1, |
| 3, |
| 512, |
| 512 |
| ] |
| }, |
| { |
| "name": "bidirectional_connector_exact", |
| "passed": true |
| }, |
| { |
| "name": "trained_query_shape", |
| "passed": true |
| }, |
| { |
| "name": "complete_28layer_SANA_flow", |
| "passed": true |
| }, |
| { |
| "name": "flow_scale_roundtrip", |
| "passed": true |
| } |
| ], |
| "smoke": { |
| "passed": true, |
| "device": "cpu", |
| "dtype": "float32" |
| }, |
| "elapsed_seconds": 41.35747504200026, |
| "limitations": [ |
| "No original Qwen language model or LM head is included; connector inputs in smoke tests are synthetic hidden states, not a text-to-image quality evaluation.", |
| "Original vision encoder, original trained custom decoder class, final six Qwen layers, projections, learned queries and full SANA DiT are loaded strictly.", |
| "The official custom decoder source is supplied externally by pinned SHA256 because its repository has no root code license.", |
| "112px/448px vision fixtures and 128px/512px RGB decode; CPU float32 tests do not establish CUDA FlashAttention numerical parity." |
| ] |
| } |
|
|
