h3-blackwell-runtime/benchmarks/gb10-sage2-vprep-canonical-repeat.json

54 lines
1.2 KiB
JSON
Raw Normal View History

{
"status": "pass",
"contract": {
"tensor_layout": "NHD",
"head_dim": 128,
"scale_max": 2.25,
"output": "Sage2 padded/permuted E4M3 V and FP32 per-channel scale"
},
"cases": [
{
"sequence": 37810,
"heads": 56,
"stride": [
813066240,
21504,
128,
1
],
"fp8": {
"equal": true,
"different_elements": 0,
"max_abs": 0.0,
"mean_abs": 0.0
},
"scale": {
"equal": true,
"different_elements": 0,
"max_abs": 0.0,
"mean_abs": 0.0
},
"baseline_timing": {
"count": 20,
"mean_s": 0.010601052856445314,
"p50_s": 0.010562719821929932,
"p90_s": 0.010675382804870605,
"p95_s": 0.010855110502243043,
"p99_s": 0.010869799871444702,
"min_s": 0.010542431831359863,
"max_s": 0.010873472213745117
},
"candidate_timing": {
"count": 20,
"mean_s": 0.006498015952110291,
"p50_s": 0.0063928160667419435,
"p90_s": 0.0067695615291595455,
"p95_s": 0.006918193387985229,
"p99_s": 0.006928189001083374,
"min_s": 0.006376319885253906,
"max_s": 0.00693068790435791
}
}
]
}