h3-blackwell-runtime/benchmarks/gb10-sage2-p0-early-k-latency-interleaved.json

199 lines
5.3 KiB
JSON
Raw Normal View History

{
"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": 0.24454946014404297,
"p50_s": 0.24409790802001952,
"p90_s": 0.24773331298828125,
"p95_s": 0.24835807189941406,
"p99_s": 0.24918752502441405,
"min_s": 0.24054057312011717,
"max_s": 0.24935398864746094
},
"rounds": [
{
"count": 10,
"mean_s": 0.24177088928222656,
"p50_s": 0.24183243560791015,
"p90_s": 0.24249975128173826,
"p95_s": 0.24257464370727538,
"p99_s": 0.24263455764770508,
"min_s": 0.24054057312011717,
"max_s": 0.2426495361328125
},
{
"count": 10,
"mean_s": 0.24227575378417968,
"p50_s": 0.2424848861694336,
"p90_s": 0.24278435516357424,
"p95_s": 0.24283492660522463,
"p99_s": 0.24287538375854495,
"min_s": 0.24105349731445314,
"max_s": 0.242885498046875
},
{
"count": 10,
"mean_s": 0.24449723815917968,
"p50_s": 0.24409790802001952,
"p90_s": 0.2455601837158203,
"p95_s": 0.24703209075927735,
"p99_s": 0.248209616394043,
"min_s": 0.24326626586914063,
"max_s": 0.24850399780273438
},
{
"count": 10,
"mean_s": 0.24652888488769534,
"p50_s": 0.24648104095458984,
"p90_s": 0.24750688781738278,
"p95_s": 0.24784330291748047,
"p99_s": 0.24811243499755858,
"min_s": 0.2454468536376953,
"max_s": 0.24817971801757813
},
{
"count": 10,
"mean_s": 0.24767453460693362,
"p50_s": 0.24752355194091796,
"p90_s": 0.24904823913574217,
"p95_s": 0.24920111389160154,
"p99_s": 0.24932341369628908,
"min_s": 0.24665737915039063,
"max_s": 0.24935398864746094
}
],
"speedup_vs_baseline_percent": 0.0
},
{
"name": "early-k",
"path": "/variants/early-k/lib/sageattention/_qattn_sm89.cpython-312-aarch64-linux-gnu.so",
"output_sha256": "4c666c20f5f8f651158a2ced33ccff08f3bada07665c595b99008d171db30574",
"byte_exact": true,
"timing": {
"count": 50,
"mean_s": 0.2447421008300781,
"p50_s": 0.2448299789428711,
"p90_s": 0.24727191772460938,
"p95_s": 0.24772567520141603,
"p99_s": 0.24888865036010743,
"min_s": 0.2410546875,
"max_s": 0.24890591430664064
},
"rounds": [
{
"count": 10,
"mean_s": 0.2419627883911133,
"p50_s": 0.2418755340576172,
"p90_s": 0.24243917236328125,
"p95_s": 0.2430596664428711,
"p99_s": 0.24355606170654298,
"min_s": 0.2410546875,
"max_s": 0.24368016052246094
},
{
"count": 10,
"mean_s": 0.24308984985351562,
"p50_s": 0.24292047882080076,
"p90_s": 0.24403975982666015,
"p95_s": 0.24468525466918944,
"p99_s": 0.24520165054321288,
"min_s": 0.24181376647949218,
"max_s": 0.24533074951171874
},
{
"count": 10,
"mean_s": 0.24468006286621097,
"p50_s": 0.24458190155029297,
"p90_s": 0.24544971313476563,
"p95_s": 0.2455271255493164,
"p99_s": 0.24558905548095702,
"min_s": 0.24351507568359376,
"max_s": 0.2456045379638672
},
{
"count": 10,
"mean_s": 0.24647027893066403,
"p50_s": 0.24653636932373046,
"p90_s": 0.24727191772460938,
"p95_s": 0.24740000457763672,
"p99_s": 0.2475024740600586,
"min_s": 0.24537295532226563,
"max_s": 0.24752809143066407
},
{
"count": 10,
"mean_s": 0.24750752410888674,
"p50_s": 0.24719758605957032,
"p90_s": 0.24887420501708982,
"p95_s": 0.24889005966186523,
"p99_s": 0.24890274337768556,
"min_s": 0.24669772338867188,
"max_s": 0.24890591430664064
}
],
"speedup_vs_baseline_percent": -0.29901196169379585
}
]
}