| { |
| "as_of": "2026-10-24", |
| "strict_load": true, |
| "source_tensor_bytes_equal": true, |
| "smoke": { |
| "passed": true, |
| "device": "cpu", |
| "runtime": "float32 EMA weights with CPU bfloat16 autocast", |
| "full_noisy_vit_layers": 24, |
| "full_decoder_layers": 16, |
| "condition_refiner_layers": 4, |
| "original_llm_or_codec_tested": false |
| }, |
| "checks": [ |
| { |
| "name": "all_selected_tensor_payload_sha256", |
| "passed": true, |
| "tensors": 662 |
| }, |
| { |
| "name": "strict_load_noisy_encoder", |
| "passed": true, |
| "tensors": 367, |
| "parameters": 566295552, |
| "checkpoint_dtype": "float32" |
| }, |
| { |
| "name": "noisy_encoder_full_depth", |
| "passed": true |
| }, |
| { |
| "name": "merged_semantic_and_deepstack_shapes", |
| "passed": true |
| }, |
| { |
| "name": "noisy_encoder_attention_parity", |
| "passed": true, |
| "relative_rms": 0.006473406683653593, |
| "max_abs": 0.015625, |
| "relative_rms_limit": 0.02 |
| }, |
| { |
| "name": "deepstack_attention_parity_0", |
| "passed": true, |
| "relative_rms": 0.004697965458035469, |
| "max_abs": 0.00390625, |
| "relative_rms_limit": 0.02 |
| }, |
| { |
| "name": "deepstack_attention_parity_1", |
| "passed": true, |
| "relative_rms": 0.005786854773759842, |
| "max_abs": 0.0078125, |
| "relative_rms_limit": 0.02 |
| }, |
| { |
| "name": "deepstack_attention_parity_2", |
| "passed": true, |
| "relative_rms": 0.005972336977720261, |
| "max_abs": 0.03125, |
| "relative_rms_limit": 0.02 |
| }, |
| { |
| "name": "time_conditioning_effect_noisy_encoder", |
| "passed": true, |
| "mean_change": 0.09781758487224579 |
| }, |
| { |
| "name": "strict_load_diffusion_decoder", |
| "passed": true, |
| "tensors": 295, |
| "parameters": 1064238144, |
| "checkpoint_dtype": "float32" |
| }, |
| { |
| "name": "decoder_full_depth", |
| "passed": true |
| }, |
| { |
| "name": "packed_patch_velocity_output", |
| "passed": true |
| }, |
| { |
| "name": "decoder_attention_parity", |
| "passed": true, |
| "relative_rms": 0.009612507186830044, |
| "max_abs": 0.01953125, |
| "relative_rms_limit": 0.02 |
| }, |
| { |
| "name": "time_conditioning_effect_diffusion_decoder", |
| "passed": true, |
| "mean_change": 0.5844433307647705 |
| }, |
| { |
| "name": "language_hidden_conditioning_effect", |
| "passed": true, |
| "mean_change": 0.39505839347839355 |
| }, |
| { |
| "name": "original_visual_position_gather_preserved", |
| "passed": true |
| }, |
| { |
| "name": "two_step_decoder_ode_fixture", |
| "passed": true, |
| "rms": [ |
| 0.3787670433521271, |
| 0.6352394223213196 |
| ], |
| "note": "Explicit Euler interface fixture; not a reproduction of the original 25-step AdamLM sampler." |
| }, |
| { |
| "name": "independent_attention_used", |
| "passed": true, |
| "attention_calls": 44 |
| } |
| ], |
| "elapsed_seconds": 155.52193593978882, |
| "limitations": [ |
| "Small packed latent and synthetic original-width LLM hidden-state fixtures, no full language/codec inference or image quality benchmark.", |
| "Native CUDA FlexAttention masks evaluated densely through SDPA for portable CPU tests; original code is retained in upstream/.", |
| "Original fixed packed modulation and condition-refiner K normalization arithmetic are preserved." |
| ] |
| } |
