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

198 lines
5.3 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": 0.24493060607910155,
"p50_s": 0.24506629180908202,
"p90_s": 0.24735857086181642,
"p95_s": 0.24807553634643553,
"p99_s": 0.2484788020324707,
"min_s": 0.24139625549316407,
"max_s": 0.24860610961914062
},
"rounds": [
{
"count": 10,
"mean_s": 0.24188405456542966,
"p50_s": 0.24180159759521486,
"p90_s": 0.2422922882080078,
"p95_s": 0.24256917114257812,
"p99_s": 0.24279067749023436,
"min_s": 0.24139625549316407,
"max_s": 0.24284605407714843
},
{
"count": 10,
"mean_s": 0.24345465087890625,
"p50_s": 0.24355528259277343,
"p90_s": 0.24430398864746095,
"p95_s": 0.24450120697021485,
"p99_s": 0.24465898162841798,
"min_s": 0.24244216918945313,
"max_s": 0.24469842529296876
},
{
"count": 10,
"mean_s": 0.24523681640625,
"p50_s": 0.24506629180908202,
"p90_s": 0.24575099334716796,
"p95_s": 0.24596200637817384,
"p99_s": 0.2461308168029785,
"min_s": 0.2445313262939453,
"max_s": 0.24617301940917968
},
{
"count": 10,
"mean_s": 0.24642633056640623,
"p50_s": 0.24648467254638673,
"p90_s": 0.24679862213134765,
"p95_s": 0.2469401741027832,
"p99_s": 0.24705341567993164,
"min_s": 0.2457111053466797,
"max_s": 0.24708172607421874
},
{
"count": 10,
"mean_s": 0.24765117797851563,
"p50_s": 0.24752259063720702,
"p90_s": 0.24837227935791015,
"p95_s": 0.24848919448852538,
"p99_s": 0.24858272659301756,
"min_s": 0.24690821838378907,
"max_s": 0.24860610961914062
}
],
"speedup_vs_baseline_percent": 0.0
},
{
"name": "scopes",
"path": "/variants/scopes/lib/sageattention/_qattn_sm89.cpython-312-aarch64-linux-gnu.so",
"output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"byte_exact": true,
"timing": {
"count": 50,
"mean_s": 0.2450569662475586,
"p50_s": 0.24491361236572265,
"p90_s": 0.2475555648803711,
"p95_s": 0.2478000213623047,
"p99_s": 0.24837403427124025,
"min_s": 0.24112467956542968,
"max_s": 0.24871961975097656
},
"rounds": [
{
"count": 10,
"mean_s": 0.24211443023681642,
"p50_s": 0.242046875,
"p90_s": 0.2430909957885742,
"p95_s": 0.24327196121215822,
"p99_s": 0.2434167335510254,
"min_s": 0.24112467956542968,
"max_s": 0.2434529266357422
},
{
"count": 10,
"mean_s": 0.24358483276367188,
"p50_s": 0.24363387298583983,
"p90_s": 0.24407558746337887,
"p95_s": 0.24423013076782224,
"p99_s": 0.24435376541137696,
"min_s": 0.24262857055664064,
"max_s": 0.24438467407226563
},
{
"count": 10,
"mean_s": 0.2451986862182617,
"p50_s": 0.24491361236572265,
"p90_s": 0.24644293365478515,
"p95_s": 0.24666307907104493,
"p99_s": 0.24683919540405272,
"min_s": 0.24451295471191406,
"max_s": 0.24688322448730468
},
{
"count": 10,
"mean_s": 0.24685755767822268,
"p50_s": 0.24679541015625,
"p90_s": 0.24760929107666016,
"p95_s": 0.24781181716918946,
"p99_s": 0.2479738380432129,
"min_s": 0.24587753295898437,
"max_s": 0.24801434326171876
},
{
"count": 10,
"mean_s": 0.2475293243408203,
"p50_s": 0.2474662857055664,
"p90_s": 0.24803896179199217,
"p95_s": 0.24837929077148435,
"p99_s": 0.24865155395507813,
"min_s": 0.24696047973632812,
"max_s": 0.24871961975097656
}
],
"speedup_vs_baseline_percent": 0.06234012143488954
}
]
}