h3-blackwell-runtime/benchmarks/gb10-h3-fused-elementwise-1344x768-124f-seed440420.json
2026-08-25 20:30:22 +07:00

77 lines
2.2 KiB
JSON

{
"device": "NVIDIA GB10",
"compute_capability": "SM121",
"torch": "2.9.1+cu130",
"attention": "sage2",
"resolution": [
1344,
768
],
"frames": 124,
"packed_tokens": 37810,
"steps": 12,
"seed": 440420,
"feature_flag": "H3_FUSED_ELEMENTWISE=1",
"isolated_cuda_tests": {
"modulation_bit_exact": true,
"gate_add_bit_exact": true
},
"real_block_results": [
{
"block": 0,
"eager_p50_seconds": 0.5083600514917634,
"fused_p50_seconds": 0.4901100995193701,
"speedup": 1.0372364331816264,
"latency_reduction": 0.03589965796651318,
"max_abs_difference": 0.0,
"checksum": -8279139.0
},
{
"block": 24,
"eager_p50_seconds": 0.5089578525221441,
"fused_p50_seconds": 0.4891547074948903,
"speedup": 1.040484420826024,
"latency_reduction": 0.03890920422804989,
"max_abs_difference": 0.0,
"checksum": 89071408.0
},
{
"block": 49,
"eager_p50_seconds": 0.5089952205016743,
"fused_p50_seconds": 0.4891957634827122,
"speedup": 1.0404734842305352,
"latency_reduction": 0.03889910203763303,
"max_abs_difference": 0.0,
"checksum": -40617144320.0
}
],
"sampling_12_step": {
"eager_seconds": 310.109976747015,
"fused_seconds": 301.04568115097936,
"speedup": 1.0301093693202321,
"latency_reduction": 0.029229293720628102,
"eager_checksums": [
-91481.7890625,
678.401611328125
],
"fused_checksums": [
-91481.7890625,
678.401611328125
]
},
"sampling_2_step_memory": {
"eager_peak_allocated_bytes": 19278355968,
"fused_peak_allocated_bytes": 18869530112,
"allocated_bytes_saved": 408825856,
"eager_peak_reserved_bytes": 21976055808,
"fused_peak_reserved_bytes": 21846032384,
"reserved_bytes_saved": 130023424
},
"validation": {
"deployed_tests": "32 passed",
"block_parity": "torch.equal at blocks 0, 24, and 49",
"sampling_parity": "identical video and audio checksums after 12 steps",
"generic_default": "eager path remains default unless H3_FUSED_ELEMENTWISE is enabled",
"spark_default": "enabled in Spark compose files"
}
}