h3-blackwell-runtime/benchmarks/gb10-sage2-p0-inplace-scores-latency-interleaved.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": 0.24575043548583986,
"p50_s": 0.24647820281982422,
"p90_s": 0.24817954711914061,
"p95_s": 0.24843196334838868,
"p99_s": 0.2488530535888672,
"min_s": 0.24133132934570312,
"max_s": 0.24893434143066406
},
"rounds": [
{
"count": 10,
"mean_s": 0.24229534759521484,
"p50_s": 0.24214422607421876,
"p90_s": 0.24312568054199218,
"p95_s": 0.24336900024414063,
"p99_s": 0.24356365600585939,
"min_s": 0.24133132934570312,
"max_s": 0.24361231994628907
},
{
"count": 10,
"mean_s": 0.24461372528076172,
"p50_s": 0.24456707000732422,
"p90_s": 0.24527583465576172,
"p95_s": 0.24591716842651368,
"p99_s": 0.24643023544311524,
"min_s": 0.2432255096435547,
"max_s": 0.24655850219726563
},
{
"count": 10,
"mean_s": 0.24647662353515623,
"p50_s": 0.24633929443359376,
"p90_s": 0.24748114624023437,
"p95_s": 0.24783390350341797,
"p99_s": 0.24811610931396483,
"min_s": 0.245321533203125,
"max_s": 0.24818666076660156
},
{
"count": 10,
"mean_s": 0.24721214752197268,
"p50_s": 0.2471304473876953,
"p90_s": 0.24781344299316405,
"p95_s": 0.24802697601318358,
"p99_s": 0.24819780242919923,
"min_s": 0.24645692443847655,
"max_s": 0.24824050903320313
},
{
"count": 10,
"mean_s": 0.24815433349609375,
"p50_s": 0.2481010208129883,
"p90_s": 0.24878503723144532,
"p95_s": 0.24885968933105468,
"p99_s": 0.2489194110107422,
"min_s": 0.2475707244873047,
"max_s": 0.24893434143066406
}
],
"speedup_vs_baseline_percent": 0.0
},
{
"name": "inplace-scores",
"path": "/variants/inplace-scores/lib/sageattention/_qattn_sm89.cpython-312-aarch64-linux-gnu.so",
"output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"byte_exact": true,
"timing": {
"count": 50,
"mean_s": 0.2458565710449219,
"p50_s": 0.2464531707763672,
"p90_s": 0.24893431243896486,
"p95_s": 0.24899965438842775,
"p99_s": 0.2501344566345215,
"min_s": 0.24153382873535156,
"max_s": 0.25087110900878906
},
"rounds": [
{
"count": 10,
"mean_s": 0.2423179229736328,
"p50_s": 0.2422579345703125,
"p90_s": 0.2428229248046875,
"p95_s": 0.24329040679931638,
"p99_s": 0.24366439239501952,
"min_s": 0.24153382873535156,
"max_s": 0.2437578887939453
},
{
"count": 10,
"mean_s": 0.2441113250732422,
"p50_s": 0.24403551483154295,
"p90_s": 0.24508229522705077,
"p95_s": 0.24537674942016602,
"p99_s": 0.2456123127746582,
"min_s": 0.24297340393066405,
"max_s": 0.24567120361328126
},
{
"count": 10,
"mean_s": 0.24651520080566405,
"p50_s": 0.24650099182128907,
"p90_s": 0.24793040618896486,
"p95_s": 0.247960213470459,
"p99_s": 0.24798405929565429,
"min_s": 0.24519821166992187,
"max_s": 0.24799002075195312
},
{
"count": 10,
"mean_s": 0.24776356201171876,
"p50_s": 0.2478148651123047,
"p90_s": 0.24900929412841796,
"p95_s": 0.2491885154724121,
"p99_s": 0.24933189254760743,
"min_s": 0.24642454528808594,
"max_s": 0.24936773681640625
},
{
"count": 10,
"mean_s": 0.24857484436035157,
"p50_s": 0.24835614776611328,
"p90_s": 0.24920902862548827,
"p95_s": 0.2500400688171387,
"p99_s": 0.250704900970459,
"min_s": 0.24763848876953126,
"max_s": 0.25087110900878906
}
],
"speedup_vs_baseline_percent": 0.010156916779835434
}
]
}