Qwen-2.5-1B-RLCD-Fast / mac_mps_initial.json
epsilon3's picture
Add tree-attention fast path and CUDA/M4 benchmarks
f36843f verified
Raw History Blame Contribute Delete
17.9 kB
{
"environment": {
"model": "Qwen/Qwen2.5-1.5B-Instruct",
"revision": "989aa7980e4cf806f80c7fef2b1adb7bc71aa306",
"torch": "2.11.0",
"transformers": "4.57.6",
"gpu": "mps",
"platform": "macOS-26.6.2-arm64-arm-64bit",
"dtype": "torch.float16",
"attention": "sdpa"
},
"arguments": {
"model": "Qwen/Qwen2.5-1.5B-Instruct",
"presets": [
"fintech_fraud",
"support_triage",
"code_security"
],
"fields": [
4,
16,
28
],
"steps": 1,
"repeats": 8,
"warmup": 2,
"output": "mac_mps_initial.json",
"device": "mps",
"dtype": "float16"
},
"cases": [
{
"preset": "fintech_fraud",
"fields": 4,
"prefix_tokens": 291,
"suffix_tokens": 28,
"batch_padded_tokens": 32,
"correctness": {
"max_abs_logit_error": 0.0390625,
"mean_abs_logit_error": 0.0043123154900968075,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": []
},
"decode": {
"batch": {
"median_ms": 47.26822950760834,
"min_ms": 46.848250000039116,
"p90_ms": 47.59587500302587,
"samples_ms": [
47.55291700712405,
47.04954198678024,
47.15175001183525,
47.52816700784024,
46.848250000039116,
47.05770799773745,
47.38470900338143,
47.59587500302587
]
},
"tree": {
"median_ms": 33.41227099735988,
"min_ms": 33.26170801301487,
"p90_ms": 34.000125000602566,
"samples_ms": [
33.38362500653602,
33.440916988183744,
33.26170801301487,
33.57716600294225,
33.26737499446608,
33.32154100644402,
33.557167000253685,
34.000125000602566
]
}
},
"end_to_end": {
"batch": {
"median_ms": 190.59108350484166,
"min_ms": 189.7072909923736,
"p90_ms": 192.90312500379514,
"samples_ms": [
191.3511250022566,
190.71195799915586,
190.8974169928115,
192.90312500379514,
190.13754100888036,
189.7072909923736,
189.88337500195485,
190.47020901052747
]
},
"tree": {
"median_ms": 178.3205000028829,
"min_ms": 177.8794160054531,
"p90_ms": 178.77554200822487,
"samples_ms": [
178.30850000609644,
177.99999999988358,
178.43274999177083,
178.77554200822487,
178.43395899399184,
177.8794160054531,
177.99320800986607,
178.33249999966938
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 3279.0,
"extra_driver_mib": 112.0
},
"tree": {
"driver_high_water_mib": 3199.0,
"extra_driver_mib": 32.0
}
},
"decode_speedup": 1.4146966996449688,
"end_to_end_speedup": 1.0688119621791121,
"decode_speedup_ci95": [
1.4070674260782117,
1.4236970070965906
],
"end_to_end_speedup_ci95": [
1.0659652991225994,
1.072772632844677
]
},
{
"preset": "fintech_fraud",
"fields": 16,
"prefix_tokens": 468,
"suffix_tokens": 127,
"batch_padded_tokens": 160,
"correctness": {
"max_abs_logit_error": 0.046875,
"mean_abs_logit_error": 0.005229064263403416,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": []
},
"decode": {
"batch": {
"median_ms": 166.3752914973884,
"min_ms": 166.01883299881592,
"p90_ms": 166.64354100066703,
"samples_ms": [
166.01883299881592,
166.21908301021904,
166.53400000359397,
166.37366599752568,
166.27183300442994,
166.3769169972511,
166.59845800313633,
166.64354100066703
]
},
"tree": {
"median_ms": 83.06970800185809,
"min_ms": 82.75195899477694,
"p90_ms": 83.72970900381915,
"samples_ms": [
82.94133399613202,
83.04516600037459,
83.72970900381915,
83.09425000334159,
83.04158299870323,
82.75195899477694,
83.13174999784678,
83.20774999447167
]
}
},
"end_to_end": {
"batch": {
"median_ms": 400.43991649872623,
"min_ms": 400.2200419927249,
"p90_ms": 402.33149999403395,
"samples_ms": [
402.33149999403395,
400.3236669959733,
400.2200419927249,
400.5150419980055,
400.364790999447,
400.3123750007944,
401.0615840088576,
401.5582089923555
]
},
"tree": {
"median_ms": 317.8602914995281,
"min_ms": 317.6447910082061,
"p90_ms": 318.26429200009443,
"samples_ms": [
317.882292001741,
317.8382909973152,
317.65129198902287,
317.7312080078991,
317.91475000500213,
317.6447910082061,
318.1417499872623,
318.26429200009443
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 5620.40625,
"extra_driver_mib": 2352.0
},
"tree": {
"driver_high_water_mib": 3276.40625,
"extra_driver_mib": 8.0
}
},
"decode_speedup": 2.00283948865773,
"end_to_end_speedup": 1.25979849389058,
"decode_speedup_ci95": [
2.0015503733169644,
2.00498794137586
],
"end_to_end_speedup_ci95": [
1.259346384504486,
1.2624706917901076
]
},
{
"preset": "fintech_fraud",
"fields": 28,
"prefix_tokens": 646,
"suffix_tokens": 227,
"batch_padded_tokens": 308,
"correctness": {
"max_abs_logit_error": 0.0390625,
"mean_abs_logit_error": 0.004572090692818165,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": [
"counterparty_jurisdiction_risk"
]
},
"decode": {
"batch": {
"median_ms": 337.68993749981746,
"min_ms": 337.26487500825897,
"p90_ms": 338.20399999967776,
"samples_ms": [
338.20399999967776,
337.63237499806564,
337.26487500825897,
337.7475000015693,
337.82583399442956,
337.49304099183064,
337.31808399898,
337.8040000097826
]
},
"tree": {
"median_ms": 142.53641650429927,
"min_ms": 142.30208400113042,
"p90_ms": 143.00833299057558,
"samples_ms": [
142.60750000539701,
142.54791699931957,
142.32374999846797,
142.30208400113042,
142.68516600714065,
143.00833299057558,
142.44258399412502,
142.52491600927897
]
}
},
"end_to_end": {
"batch": {
"median_ms": 665.5089580017375,
"min_ms": 664.9612079927465,
"p90_ms": 671.5790419984842,
"samples_ms": [
671.5790419984842,
665.4760410019662,
666.8390420090873,
665.4567090008641,
664.9612079927465,
665.394916999503,
665.5418750015087,
668.8652919983724
]
},
"tree": {
"median_ms": 471.9899789997726,
"min_ms": 470.15170801023487,
"p90_ms": 473.9503750024596,
"samples_ms": [
472.25875000003725,
473.07704100967385,
471.5012079977896,
471.7212079995079,
472.3985000018729,
470.7340419990942,
470.15170801023487,
473.9503750024596
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 5294.984375,
"extra_driver_mib": 2022.0
},
"tree": {
"driver_high_water_mib": 3270.984375,
"extra_driver_mib": 40.0
}
},
"decode_speedup": 2.369148501005228,
"end_to_end_speedup": 1.4100065416898568,
"decode_speedup_ci95": [
2.366431024764677,
2.3721027360653832
],
"end_to_end_speedup_ci95": [
1.4077065183304176,
1.4155099921366008
]
},
{
"preset": "support_triage",
"fields": 4,
"prefix_tokens": 298,
"suffix_tokens": 28,
"batch_padded_tokens": 40,
"correctness": {
"max_abs_logit_error": 0.0390625,
"mean_abs_logit_error": 0.004746712744235992,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": []
},
"decode": {
"batch": {
"median_ms": 51.41572900902247,
"min_ms": 50.98366699530743,
"p90_ms": 52.308957994682714,
"samples_ms": [
51.25137499999255,
50.98366699530743,
51.47558300814126,
51.355875009903684,
52.004624987603165,
52.308957994682714,
51.34412499319296,
51.97441601194441
]
},
"tree": {
"median_ms": 33.48941699368879,
"min_ms": 33.176375000039116,
"p90_ms": 35.18329199869186,
"samples_ms": [
33.21337500528898,
33.488749992102385,
33.33837499667425,
33.89491600682959,
35.18329199869186,
33.57179199520033,
33.176375000039116,
33.4900839952752
]
}
},
"end_to_end": {
"batch": {
"median_ms": 194.8159375024261,
"min_ms": 194.30545800423715,
"p90_ms": 196.05383300222456,
"samples_ms": [
194.53891599550843,
194.30545800423715,
194.7874170000432,
194.8301250085933,
196.05383300222456,
195.8960420015501,
195.53312499192543,
194.80174999625888
]
},
"tree": {
"median_ms": 179.28256250161212,
"min_ms": 178.82008400920313,
"p90_ms": 181.21854099445045,
"samples_ms": [
179.84216600598302,
179.1274169954704,
178.82008400920313,
178.8337500038324,
180.26720799389295,
181.21854099445045,
179.43770800775383,
179.10737499187235
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 3359.15625,
"extra_driver_mib": 104.0
},
"tree": {
"driver_high_water_mib": 3263.15625,
"extra_driver_mib": 8.0
}
},
"decode_speedup": 1.5352828930617681,
"end_to_end_speedup": 1.0866418617854947,
"decode_speedup_ci95": [
1.5224064292082697,
1.5519344776584314
],
"end_to_end_speedup_ci95": [
1.0831418894925375,
1.0896991895565011
]
},
{
"preset": "support_triage",
"fields": 16,
"prefix_tokens": 481,
"suffix_tokens": 126,
"batch_padded_tokens": 160,
"correctness": {
"max_abs_logit_error": 0.046875,
"mean_abs_logit_error": 0.0048467740416526794,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": []
},
"decode": {
"batch": {
"median_ms": 168.7676039946382,
"min_ms": 168.39800000889227,
"p90_ms": 169.3617080018157,
"samples_ms": [
168.98479100200348,
168.74533299414907,
169.21074999845587,
168.39800000889227,
168.68324999813922,
168.61854199669324,
168.78987499512732,
169.3617080018157
]
},
"tree": {
"median_ms": 83.00702099950286,
"min_ms": 82.86941600090358,
"p90_ms": 83.5149589984212,
"samples_ms": [
83.5149589984212,
82.96704200620297,
83.02208301029168,
82.99604199419264,
82.9257090081228,
82.86941600090358,
83.01800000481308,
83.453457991709
]
}
},
"end_to_end": {
"batch": {
"median_ms": 408.5541460008244,
"min_ms": 407.3758330050623,
"p90_ms": 411.9953330082353,
"samples_ms": [
409.455084009096,
408.30883300805,
407.3758330050623,
407.4040419945959,
408.7994589935988,
407.83354100130964,
408.8325830089161,
411.9953330082353
]
},
"tree": {
"median_ms": 322.66933350183535,
"min_ms": 321.90650000120513,
"p90_ms": 325.09766600560397,
"samples_ms": [
322.21112499246374,
325.09766600560397,
322.2837920038728,
321.90650000120513,
323.0462500068825,
322.482667004806,
323.7660419981694,
322.8559999988647
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 5599.484375,
"extra_driver_mib": 2352.0
},
"tree": {
"driver_high_water_mib": 3271.484375,
"extra_driver_mib": 24.0
}
},
"decode_speedup": 2.033172639645133,
"end_to_end_speedup": 1.2661697396740696,
"decode_speedup_ci95": [
2.0271568557257,
2.036780629549704
],
"end_to_end_speedup_ci95": [
1.2619319354399532,
1.2698004562026703
]
},
{
"preset": "support_triage",
"fields": 28,
"prefix_tokens": 655,
"suffix_tokens": 218,
"batch_padded_tokens": 280,
"correctness": {
"max_abs_logit_error": 0.046875,
"mean_abs_logit_error": 0.004553103819489479,
"full_vocab_argmax_agreement": 1.0,
"candidate_decisions_equal": true,
"candidate_collision_fields": [
"target_resolution_hours",
"assigned_agent_tier"
]
},
"decode": {
"batch": {
"median_ms": 333.23189599468606,
"min_ms": 332.5708750053309,
"p90_ms": 338.84008300083224,
"samples_ms": [
333.4462920029182,
332.5708750053309,
333.0174999864539,
333.4563749958761,
338.84008300083224,
333.65920900541823,
332.6203750038985,
332.86400001088623
]
},
"tree": {
"median_ms": 140.46581200091168,
"min_ms": 140.3020409925375,
"p90_ms": 141.14466699538752,
"samples_ms": [
141.14466699538752,
140.44829200429376,
140.45941599761136,
140.65008399484213,
141.10125000297558,
140.3020409925375,
140.472208004212,
140.3820419945987
]
}
},
"end_to_end": {
"batch": {
"median_ms": 664.504312997451,
"min_ms": 662.5099170050817,
"p90_ms": 666.6135840059724,
"samples_ms": [
664.545999999973,
664.5135419967119,
662.5099170050817,
664.9717500113184,
664.3663330032723,
666.6135840059724,
664.3937079934403,
664.4950839981902
]
},
"tree": {
"median_ms": 471.8639579950832,
"min_ms": 470.3940409963252,
"p90_ms": 473.91804099606816,
"samples_ms": [
473.1918329925975,
470.3940409963252,
473.91804099606816,
471.0846249945462,
471.2021250015823,
472.3992079962045,
472.50158300448675,
471.32870799396187
]
}
},
"memory": {
"batch": {
"driver_high_water_mib": 5292.015625,
"extra_driver_mib": 2062.0
},
"tree": {
"driver_high_water_mib": 3276.015625,
"extra_driver_mib": 48.0
}
},
"decode_speedup": 2.372334529291179,
"end_to_end_speedup": 1.408254014188503,
"decode_speedup_ci95": [
2.366582759337612,
2.376035750477272
],
"end_to_end_speedup_ci95": [
1.4043364565201133,
1.4109481535389878
]
}
]
}