internvlu_image_dit / validation.json
gatilin's picture
Publish verified native DiT using pinned official LFS copy and numerical validation
a130d59 verified
Raw History Blame Contribute Delete
3.24 kB
{
"as_of": "2026-10-28",
"strict_load": true,
"source_tensor_bytes_equal": true,
"smoke": {
"passed": true,
"device": "cpu",
"runtime_dtype": "float32",
"synthetic_conditions": true,
"real_text_or_audio_conditioning_tested": false,
"codec_decoding_tested": false
},
"checks": [
{
"name": "export_weight_sha256",
"passed": true
},
{
"name": "strict_native_load_all_keys_and_shapes",
"passed": true,
"tensors": 654,
"parameters": 1808034880,
"original_dtypes": {
"BF16": 1713511936,
"F32": 94522944
}
},
{
"name": "nonlearned_rope_cache_materialized",
"passed": true
},
{
"name": "native_complete_20_layers",
"passed": true
},
{
"name": "rectangular_native_latent_forward",
"passed": true,
"rms": 0.5196120142936707
},
{
"name": "original_native_vs_exported_interface",
"passed": true,
"max_abs": 0.0,
"atol": 0,
"rtol": 0
},
{
"name": "all_layers_sdpa_vs_explicit_attention",
"passed": true,
"max_abs": 1.9669532775878906e-06,
"atol": 0.0005,
"rtol": 0.0005
},
{
"name": "conditioning_changes_prediction",
"passed": true
},
{
"name": "time_changes_prediction",
"passed": true
},
{
"name": "masked_condition_isolation",
"passed": true,
"max_abs": 0.0,
"atol": 0.0003,
"rtol": 0.0003
},
{
"name": "batch_isolation_first",
"passed": true,
"max_abs": 4.582107067108154e-06,
"atol": 0.0005,
"rtol": 0.0005
},
{
"name": "batch_isolation_second",
"passed": true,
"max_abs": 4.947185516357422e-06,
"atol": 0.0005,
"rtol": 0.0005
},
{
"name": "semantic_visual_condition_grid",
"passed": true
},
{
"name": "reference_vae_editing_input",
"passed": true
},
{
"name": "native_dpm_flow_scheduler",
"passed": true
},
{
"name": "three_step_native_dpm_trajectory",
"passed": true,
"latent_rms": [
0.15130993723869324,
0.14316070079803467,
0.3429756164550781
]
},
{
"name": "independent_attention_reference_exercised",
"passed": true,
"calls": 20
}
],
"tensors": 654,
"parameters": 1808034880,
"original_dtype_elements": {
"BF16": 1713511936,
"F32": 94522944
},
"elapsed_seconds": 21.923,
"versions": {
"torch": "2.14.1",
"safetensors": "0.8.0",
"diffusers": "0.35.2",
"transformers": "4.57.6",
"einops": "0.8.2",
"huggingface_hub": "0.36.2"
},
"limitations": [
"Full-depth real-weight CPU component fixtures using synthetic conditions and native latents.",
"No full VLM/Talker conditioning or image/audio codec decoding; no end-to-end generation-quality claim.",
"Checkpoint dtypes are preserved; float32 is a runtime validation cast only."
],
"timezone": "Asia/Shanghai",
"loader_repair": "Reconstruct the upstream deterministic RoPE cache outside meta initialization; trained weights and native computational source are unchanged."
}