45 lines
1.5 KiB
JSON
45 lines
1.5 KiB
JSON
{
|
|
"device": "NVIDIA GB10",
|
|
"compute_capability": "SM121",
|
|
"workload": {
|
|
"resolution": [1344, 768],
|
|
"frames": 124,
|
|
"packed_tokens": 37810,
|
|
"seed": 440420,
|
|
"attention": "sage2"
|
|
},
|
|
"feature_flag": "H3_NVFP4_MODULATE_FUSION=1",
|
|
"implementation": "Exact BF16 AdaLN modulation recomputed in Vortex global-scale and native QDATA/SFA pack passes; Comfy NVFP4 GEMMs unchanged",
|
|
"producer_validation": {
|
|
"blocks": [0, 24, 49],
|
|
"rows": 37810,
|
|
"tensor_scale_equal": true,
|
|
"qdata_differences": 0,
|
|
"sfa_differences": 0,
|
|
"block_24_producer_p50_ms": 6.018611951731145
|
|
},
|
|
"alternating_block_p50_improvement_percent": {
|
|
"block_0": 0.2794895723218782,
|
|
"block_24": 0.6017027236613792,
|
|
"block_49": 0.7939711616055223
|
|
},
|
|
"warmed_trajectory": {
|
|
"two_step": {
|
|
"baseline_seconds": 47.66834012803156,
|
|
"candidate_seconds": 47.41923755296739,
|
|
"improvement_percent": 0.5225744684944189,
|
|
"video_equal": true,
|
|
"audio_equal": true
|
|
},
|
|
"twelve_step": {
|
|
"baseline_seconds": 285.99590051098494,
|
|
"candidate_seconds": 284.4078275830252,
|
|
"improvement_percent": 0.5552782138213641,
|
|
"video_checksum": -91481.7890625,
|
|
"audio_checksum": 678.401611328125,
|
|
"video_equal": true,
|
|
"audio_equal": true
|
|
}
|
|
},
|
|
"decision": "Enable for Spark single-GPU deployments. Distributed, chunked-MLP, full-precision, pre-scale, autograd, and active-LoRA cases retain materialized fallback."
|
|
}
|