h3-blackwell-runtime/benchmarks/gb10-sage2-p0-interleave-softmax-latency.json
2026-08-25 20:30:22 +07:00

198 lines
5.4 KiB
JSON

{
"metadata": {
"width": 1344,
"height": 768,
"frames": 124,
"steps": 12,
"sampler_step": 1,
"seed": 440420,
"text_tokens": 100,
"block_index": 24,
"attention": "sage2",
"hidden_shape": [
37810,
5376
],
"h_msa_shape": [
37810,
5376
],
"segments": [
[
0,
100,
1
],
[
100,
514,
2
],
[
514,
37810,
0
]
]
},
"source_commit": "d1a57a546c3d395b1ffcbeecc66d81db76f3b4b5",
"reference_output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"q_shape": [
1,
37810,
56,
128
],
"scheduler": {
"cta_q": 128,
"cta_k": 64,
"warp_q": 32,
"warp_k": 64,
"threads_per_cta": 128,
"dynamic_shared_memory_bytes": 32768,
"grid_ctas": 16576
},
"rounds": 5,
"warmup": 3,
"iterations_per_round": 10,
"variants": [
{
"name": "baseline",
"path": "/variants/baseline/lib/sageattention/_qattn_sm89.cpython-312-aarch64-linux-gnu.so",
"output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"byte_exact": true,
"timing": {
"count": 50,
"mean_s": 0.24388563781738282,
"p50_s": 0.24391904449462892,
"p90_s": 0.24706241455078126,
"p95_s": 0.2474617546081543,
"p99_s": 0.24945008117675782,
"min_s": 0.23941552734375,
"max_s": 0.250071044921875
},
"rounds": [
{
"count": 10,
"mean_s": 0.24110272827148438,
"p50_s": 0.24101924896240234,
"p90_s": 0.24182757263183594,
"p95_s": 0.24287572326660156,
"p99_s": 0.24371424377441406,
"min_s": 0.23941552734375,
"max_s": 0.2439238739013672
},
{
"count": 10,
"mean_s": 0.24202915649414064,
"p50_s": 0.2419222717285156,
"p90_s": 0.24283133544921875,
"p95_s": 0.24342016143798828,
"p99_s": 0.2438912222290039,
"min_s": 0.24108163452148437,
"max_s": 0.2440089874267578
},
{
"count": 10,
"mean_s": 0.24356186828613283,
"p50_s": 0.2436042709350586,
"p90_s": 0.24417928314208984,
"p95_s": 0.24448768997192383,
"p99_s": 0.244734415435791,
"min_s": 0.2425731201171875,
"max_s": 0.24479609680175782
},
{
"count": 10,
"mean_s": 0.24519928436279298,
"p50_s": 0.24513208007812498,
"p90_s": 0.24575088195800782,
"p95_s": 0.2460108322143555,
"p99_s": 0.24621879241943362,
"min_s": 0.24462042236328124,
"max_s": 0.24627078247070314
},
{
"count": 10,
"mean_s": 0.24753515167236326,
"p50_s": 0.24709844970703126,
"p90_s": 0.2489304992675781,
"p95_s": 0.24950077209472654,
"p99_s": 0.24995699035644528,
"min_s": 0.24651388549804687,
"max_s": 0.250071044921875
}
],
"speedup_vs_baseline_percent": 0.0
},
{
"name": "interleave-softmax",
"path": "/variants/interleave-softmax/lib/sageattention/_qattn_sm89.cpython-312-aarch64-linux-gnu.so",
"output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"byte_exact": true,
"timing": {
"count": 50,
"mean_s": 0.24397127563476562,
"p50_s": 0.2439577865600586,
"p90_s": 0.24716617736816404,
"p95_s": 0.2477003547668457,
"p99_s": 0.2484949363708496,
"min_s": 0.239262939453125,
"max_s": 0.24906739807128905
},
"rounds": [
{
"count": 10,
"mean_s": 0.24102854003906252,
"p50_s": 0.24122242736816407,
"p90_s": 0.24205140991210936,
"p95_s": 0.24227293548583984,
"p99_s": 0.24245015594482422,
"min_s": 0.239262939453125,
"max_s": 0.24249446105957032
},
{
"count": 10,
"mean_s": 0.24219576721191408,
"p50_s": 0.24199546813964845,
"p90_s": 0.24289817504882813,
"p95_s": 0.24381009521484376,
"p99_s": 0.24453963134765627,
"min_s": 0.24077565002441406,
"max_s": 0.2447220153808594
},
{
"count": 10,
"mean_s": 0.2438770950317383,
"p50_s": 0.2437736587524414,
"p90_s": 0.24470800323486327,
"p95_s": 0.24506654891967772,
"p99_s": 0.2453533854675293,
"min_s": 0.24293548583984376,
"max_s": 0.2454250946044922
},
{
"count": 10,
"mean_s": 0.24551427154541017,
"p50_s": 0.2453832778930664,
"p90_s": 0.2465328369140625,
"p95_s": 0.24662471008300782,
"p99_s": 0.24669820861816405,
"min_s": 0.24450390625,
"max_s": 0.24671658325195311
},
{
"count": 10,
"mean_s": 0.24724070434570314,
"p50_s": 0.24736494445800783,
"p90_s": 0.24801593780517578,
"p95_s": 0.2485416679382324,
"p99_s": 0.24896225204467773,
"min_s": 0.24577845764160156,
"max_s": 0.24906739807128905
}
],
"speedup_vs_baseline_percent": -0.015880643112875514
}
]
}