Document Frontier fixed-cohort fidelity results

This commit is contained in:
2026-07-16 06:21:40 +08:00
parent 344af3a428
commit 13f0713bf2
9 changed files with 19154 additions and 0 deletions

View File

@@ -0,0 +1,790 @@
{
"cohort": [
{
"input_bin": 0,
"input_tokens": 865,
"prompt_sha256": "c9e22fa0559b409b259bdb15ed6eec1bb36ecb8611ae367b2e122bc6303f56ae",
"sampling_u": 0.46149221772182225,
"source_arrival_s": 8.92599999999993,
"source_request_id": "250",
"source_row_index": 250
},
{
"input_bin": 4,
"input_tokens": 9154,
"prompt_sha256": "53b0a0af6cb1c0278183c61b8b6c026977b6cc46e07fca825b5bc8c5802b0e17",
"sampling_u": 0.8156596361943489,
"source_arrival_s": 12.708999999999833,
"source_request_id": "367",
"source_row_index": 367
},
{
"input_bin": 0,
"input_tokens": 718,
"prompt_sha256": "87d2846513ce3def0c4a7f9d3707f2eda02069f74535ba69dba6fb761bbfadf5",
"sampling_u": 0.9701010298706338,
"source_arrival_s": 25.190000000000055,
"source_request_id": "682",
"source_row_index": 682
},
{
"input_bin": 1,
"input_tokens": 1519,
"prompt_sha256": "7dbffc44084f1dfd9af43c2c40b5e669728a2c01b0a1e29f65ea17bd37e5bffc",
"sampling_u": 0.48115960917218603,
"source_arrival_s": 25.639999999999873,
"source_request_id": "690",
"source_row_index": 690
},
{
"input_bin": 1,
"input_tokens": 1968,
"prompt_sha256": "65e1015fbc28a8a2af5ffced6947478c4d68b39cd07a79c5e2fb937ecf9d64a7",
"sampling_u": 0.7394708415743882,
"source_arrival_s": 36.82400000000007,
"source_request_id": "1022",
"source_row_index": 1022
},
{
"input_bin": 0,
"input_tokens": 62,
"prompt_sha256": "6ba251e2a1653df7c13c70ed9aa3317af23f42d0ec14b14e46b1376b86f548b7",
"sampling_u": 0.5901445515787429,
"source_arrival_s": 39.96399999999994,
"source_request_id": "1098",
"source_row_index": 1098
},
{
"input_bin": 0,
"input_tokens": 702,
"prompt_sha256": "ec3dbd6ac6674268fb420b0e006f7edbbad38906f178faf4c41d5ac3aa7a7968",
"sampling_u": 0.4229843993649856,
"source_arrival_s": 60.55600000000004,
"source_request_id": "1571",
"source_row_index": 1571
},
{
"input_bin": 1,
"input_tokens": 1472,
"prompt_sha256": "1fc79d848cc4d3f92ee5a9ad6ecc28a3173737f77ec0b2eaca1b38026225c701",
"sampling_u": 0.38189150492188706,
"source_arrival_s": 69.46300000000019,
"source_request_id": "1792",
"source_row_index": 1792
},
{
"input_bin": 2,
"input_tokens": 2701,
"prompt_sha256": "084b29ad36178a06d93a1fdfbed02bc0718b42f43187d184bc06f70f03df4dfc",
"sampling_u": 0.8169083216699623,
"source_arrival_s": 69.79100000000017,
"source_request_id": "1799",
"source_row_index": 1799
},
{
"input_bin": 3,
"input_tokens": 7825,
"prompt_sha256": "044f618323ac49afa864bf6b62b1e06a9f3c351647a953cbe38b3173542caf51",
"sampling_u": 0.3751377474987089,
"source_arrival_s": 70.71099999999979,
"source_request_id": "1818",
"source_row_index": 1818
},
{
"input_bin": 0,
"input_tokens": 715,
"prompt_sha256": "e5f9aba41593f81170000f1299c71d0f09f52afc58addc2968fb862f8b740359",
"sampling_u": 0.7098663141194472,
"source_arrival_s": 75.24699999999984,
"source_request_id": "1935",
"source_row_index": 1935
},
{
"input_bin": 1,
"input_tokens": 1573,
"prompt_sha256": "d69852bb4dbc6b7653ed3a200e891190af5d052d20e57c2ff3abbccc6bf254aa",
"sampling_u": 0.6074460889792176,
"source_arrival_s": 84.5329999999999,
"source_request_id": "2176",
"source_row_index": 2176
},
{
"input_bin": 5,
"input_tokens": 29385,
"prompt_sha256": "86cb5ce843bad59ada0e7b9e1a60c1162534b037f5c1b5a7fadca5be26254047",
"sampling_u": 0.17186958800332738,
"source_arrival_s": 93.04399999999987,
"source_request_id": "2396",
"source_row_index": 2396
},
{
"input_bin": 1,
"input_tokens": 1939,
"prompt_sha256": "06bad3b6fc6db1c9227a8bb2d4ce349b31802e51e056b541858773debb0a388d",
"sampling_u": 0.29421530368234206,
"source_arrival_s": 98.4050000000002,
"source_request_id": "2505",
"source_row_index": 2505
},
{
"input_bin": 1,
"input_tokens": 1532,
"prompt_sha256": "9117e580cf5b8dcedb3c6c02318610c4780fbbf1c4413367db6f7eb60b73459b",
"sampling_u": 0.41616945652066173,
"source_arrival_s": 116.05200000000013,
"source_request_id": "2913",
"source_row_index": 2913
},
{
"input_bin": 0,
"input_tokens": 513,
"prompt_sha256": "6cf0a32f0a62d4a16f6abddf0bed8af587ba8f0aff335fd54c72f5772da4464a",
"sampling_u": 0.3613591846324354,
"source_arrival_s": 118.6239999999998,
"source_request_id": "2979",
"source_row_index": 2979
},
{
"input_bin": 1,
"input_tokens": 1508,
"prompt_sha256": "ad49c5bb6cf0ced603cc052215c18159d5c25927048ddb7efd6bf315bd84c937",
"sampling_u": 0.1650938738573754,
"source_arrival_s": 121.26600000000008,
"source_request_id": "3056",
"source_row_index": 3056
},
{
"input_bin": 0,
"input_tokens": 705,
"prompt_sha256": "c7b6f66c2ade7054abed59f4459bbe4ab4659a3d53654197d27ce26f650450e1",
"sampling_u": 0.8484961251249794,
"source_arrival_s": 130.1909999999998,
"source_request_id": "3301",
"source_row_index": 3301
},
{
"input_bin": 4,
"input_tokens": 8287,
"prompt_sha256": "e597785700703dc05da72b1045e734f326ed491276f2d9f45f663f03f1ccd509",
"sampling_u": 0.7289458530141959,
"source_arrival_s": 133.10300000000007,
"source_request_id": "3379",
"source_row_index": 3379
},
{
"input_bin": 1,
"input_tokens": 1756,
"prompt_sha256": "b9631ff1dfec71fbd9e3a084bccf30e40a058b74743ae9878e3dd4be77efdf83",
"sampling_u": 0.4082222424943361,
"source_arrival_s": 152.64800000000014,
"source_request_id": "3844",
"source_row_index": 3844
},
{
"input_bin": 0,
"input_tokens": 703,
"prompt_sha256": "a943002270d455f4e65022747cf40807044f47fb574089d95deadffed56fcf2e",
"sampling_u": 0.49508596140129646,
"source_arrival_s": 166.43499999999995,
"source_request_id": "4163",
"source_row_index": 4163
},
{
"input_bin": 1,
"input_tokens": 1478,
"prompt_sha256": "4b84065a6336bfb9ddd174099740689adfeb5ce294fe670885334934cf583d8b",
"sampling_u": 0.2476209159147146,
"source_arrival_s": 172.2800000000002,
"source_request_id": "4323",
"source_row_index": 4323
},
{
"input_bin": 0,
"input_tokens": 686,
"prompt_sha256": "b2154c41a891d72e417d3bb3d67920590349860655d93cb3ff478ad3da74e2f8",
"sampling_u": 0.7190432102580353,
"source_arrival_s": 193.67599999999993,
"source_request_id": "4838",
"source_row_index": 4838
},
{
"input_bin": 1,
"input_tokens": 1600,
"prompt_sha256": "9f4cafecff466decfe98d4a9544d05c6246853f24cfeb6e77c9e4d1a5cb6cb02",
"sampling_u": 0.71711230795038,
"source_arrival_s": 194.66400000000021,
"source_request_id": "4870",
"source_row_index": 4870
},
{
"input_bin": 2,
"input_tokens": 4000,
"prompt_sha256": "30c385ba8463097d52f838bbfa69f18a32c91b38878e15c11e6b69d6bf81abc7",
"sampling_u": 0.12696475188087025,
"source_arrival_s": 216.49299999999994,
"source_request_id": "5469",
"source_row_index": 5469
},
{
"input_bin": 0,
"input_tokens": 710,
"prompt_sha256": "807c8b29c95a0623c585ee7d15d82e72be5858f3683d3657ef3295e71ef6aa10",
"sampling_u": 0.32798737479046675,
"source_arrival_s": 223.39199999999983,
"source_request_id": "5648",
"source_row_index": 5648
},
{
"input_bin": 0,
"input_tokens": 727,
"prompt_sha256": "b872bc3174db8e4b17e182608dbb48f99196220dfd55096c2490484a6cadd595",
"sampling_u": 0.3644506007131907,
"source_arrival_s": 239.0999999999999,
"source_request_id": "6031",
"source_row_index": 6031
},
{
"input_bin": 4,
"input_tokens": 8786,
"prompt_sha256": "5757f2f9e6be7d4a29a79ad82408e51069044828f55b8c21b7216abbf73b0757",
"sampling_u": 0.12629800224853702,
"source_arrival_s": 260.4920000000002,
"source_request_id": "6604",
"source_row_index": 6604
},
{
"input_bin": 4,
"input_tokens": 9518,
"prompt_sha256": "730aee2b4ec388efbf4ca55cbfcbe31c8aa9636c8876b624b37b499d5dcb00f5",
"sampling_u": 0.29338926296423934,
"source_arrival_s": 260.91499999999996,
"source_request_id": "6611",
"source_row_index": 6611
},
{
"input_bin": 0,
"input_tokens": 796,
"prompt_sha256": "e0f7312dcc97af6236edac6e47a828f9e18e3e118023537d82f6cb540025191e",
"sampling_u": 0.47864598488768767,
"source_arrival_s": 274.52,
"source_request_id": "6905",
"source_row_index": 6905
},
{
"input_bin": 0,
"input_tokens": 228,
"prompt_sha256": "b85c14bcfaec63a99ebc382a7e20d32390a65999062e84c303441344832c3279",
"sampling_u": 0.8157098758872824,
"source_arrival_s": 274.6399999999999,
"source_request_id": "6911",
"source_row_index": 6911
},
{
"input_bin": 5,
"input_tokens": 18439,
"prompt_sha256": "8f7259429d8b3b6ab062c885d990fa2136e8ac061ce5b59169e5dbe7cdd5c055",
"sampling_u": 0.44219574639241527,
"source_arrival_s": 287.1959999999999,
"source_request_id": "7271",
"source_row_index": 7271
},
{
"input_bin": 0,
"input_tokens": 736,
"prompt_sha256": "1c56453602e159f1dfaae09b5b40785d0dbc176d9b27844362e342e3731c2032",
"sampling_u": 0.6889130017060773,
"source_arrival_s": 291.97400000000016,
"source_request_id": "7403",
"source_row_index": 7403
},
{
"input_bin": 4,
"input_tokens": 8557,
"prompt_sha256": "14346a1cc9d1382dd67eaf700f5b52798b77fc8a08bc1df71768d38d613e9e43",
"sampling_u": 0.3112632330750767,
"source_arrival_s": 292.692,
"source_request_id": "7422",
"source_row_index": 7422
},
{
"input_bin": 1,
"input_tokens": 1586,
"prompt_sha256": "4d24bf64da9142747d8ee8d87daf50f81309928c3d4792f7c345ecd9dd0c24ac",
"sampling_u": 0.23031083293934718,
"source_arrival_s": 304.69900000000007,
"source_request_id": "7741",
"source_row_index": 7741
},
{
"input_bin": 1,
"input_tokens": 1871,
"prompt_sha256": "5a4917df4ebb8a2e4e3dcb0f15c231cd46830af3dc63452ebf421b4433756159",
"sampling_u": 0.10577279384258167,
"source_arrival_s": 312.683,
"source_request_id": "7981",
"source_row_index": 7981
},
{
"input_bin": 0,
"input_tokens": 698,
"prompt_sha256": "610618e93c2f4d7bd0ca26b2611b61ffdbbbe092fe0ff3a5f7d7c91f52de7c74",
"sampling_u": 0.43535389643319383,
"source_arrival_s": 319.1579999999999,
"source_request_id": "8179",
"source_row_index": 8179
},
{
"input_bin": 0,
"input_tokens": 710,
"prompt_sha256": "3b397d03d7a7001ab8f3915b1b88ed8f8b859ad51769ac752ff35f567ccb1183",
"sampling_u": 0.6383291776591391,
"source_arrival_s": 321.4380000000001,
"source_request_id": "8224",
"source_row_index": 8224
},
{
"input_bin": 1,
"input_tokens": 1452,
"prompt_sha256": "671c93bcdd7a292d1a4b4dcce44c679369955fdb4d250d451170c63391216436",
"sampling_u": 0.12477813551209259,
"source_arrival_s": 337.8319999999999,
"source_request_id": "8612",
"source_row_index": 8612
},
{
"input_bin": 5,
"input_tokens": 24271,
"prompt_sha256": "a84a31f74c755d02154a7eaeec8bd71450c16813d8b7e713999ff52419f84ed6",
"sampling_u": 0.5538066537263959,
"source_arrival_s": 348.692,
"source_request_id": "8891",
"source_row_index": 8891
},
{
"input_bin": 3,
"input_tokens": 4607,
"prompt_sha256": "720cf238e668f4ae567ab909b2ec0185d222a14eb474ae2ef481141a885b3aad",
"sampling_u": 0.03539900309623126,
"source_arrival_s": 382.55499999999984,
"source_request_id": "9768",
"source_row_index": 9768
},
{
"input_bin": 0,
"input_tokens": 111,
"prompt_sha256": "8a12215d12ddcf36d1e9013ec0b4a433e4f27c5c2a774ead0f66c93bb31a8063",
"sampling_u": 0.9241501254905812,
"source_arrival_s": 402.8620000000001,
"source_request_id": "10305",
"source_row_index": 10305
},
{
"input_bin": 0,
"input_tokens": 714,
"prompt_sha256": "e1dd5099fb8565866691241945b15ee40a5ea8ad9045d10e1dd7774b668d9fa4",
"sampling_u": 0.4533435244095546,
"source_arrival_s": 405.0459999999998,
"source_request_id": "10383",
"source_row_index": 10383
},
{
"input_bin": 1,
"input_tokens": 1792,
"prompt_sha256": "9d94bba8c056a775a5d00ecfd6567e83030c99dc20f71ce53d787683ca2450e9",
"sampling_u": 0.8983379345538031,
"source_arrival_s": 405.08100000000013,
"source_request_id": "10385",
"source_row_index": 10385
},
{
"input_bin": 0,
"input_tokens": 697,
"prompt_sha256": "7a66c3cec2e1727fa78ed1a7dbe1d39d2c877cc2835c0b10479b396204749984",
"sampling_u": 0.43645445582603326,
"source_arrival_s": 413.16499999999996,
"source_request_id": "10618",
"source_row_index": 10618
},
{
"input_bin": 0,
"input_tokens": 709,
"prompt_sha256": "a71447eebb513ece438e0a81a0a9788fbbf34a71d9a66987e6f476cf76f0b7d5",
"sampling_u": 0.4730229506348141,
"source_arrival_s": 424.991,
"source_request_id": "10911",
"source_row_index": 10911
},
{
"input_bin": 0,
"input_tokens": 699,
"prompt_sha256": "e04aaa160224e82157f9b0c89180dc341e2c8eebbd518ea8d3327eb73ebfac1c",
"sampling_u": 0.5350250700640652,
"source_arrival_s": 430.51800000000003,
"source_request_id": "11058",
"source_row_index": 11058
},
{
"input_bin": 2,
"input_tokens": 3130,
"prompt_sha256": "ba93e2babcf7a0491ba97930f5b3bfb3cc2170dc3cd7b3ca3c3f9fe2905dc72e",
"sampling_u": 0.6597844956490088,
"source_arrival_s": 434.3449999999998,
"source_request_id": "11166",
"source_row_index": 11166
},
{
"input_bin": 0,
"input_tokens": 711,
"prompt_sha256": "80c05739d748df64d97f92742f28e384365b26f4f5f1e999666ea2f7f187fd16",
"sampling_u": 0.7240395487416784,
"source_arrival_s": 441.3699999999999,
"source_request_id": "11373",
"source_row_index": 11373
},
{
"input_bin": 1,
"input_tokens": 1536,
"prompt_sha256": "ee1209838aca67d2392bb9156b74d69de22d166d97c88b56501ca9cfcc70578e",
"sampling_u": 0.9806533325226275,
"source_arrival_s": 463.8670000000002,
"source_request_id": "11911",
"source_row_index": 11911
},
{
"input_bin": 1,
"input_tokens": 1580,
"prompt_sha256": "a4e5fdb5d6119e9207384fd3d32ac027469b3ecca278738c7078026551999786",
"sampling_u": 0.8508070546796023,
"source_arrival_s": 471.08899999999994,
"source_request_id": "12100",
"source_row_index": 12100
},
{
"input_bin": 1,
"input_tokens": 1466,
"prompt_sha256": "09a6942491467ee3597eef41c4f36de46bd88b0533af8c1084bb9e15cb150832",
"sampling_u": 0.03536418626599094,
"source_arrival_s": 488.5859999999998,
"source_request_id": "12587",
"source_row_index": 12587
},
{
"input_bin": 0,
"input_tokens": 750,
"prompt_sha256": "6a53081dfa161f2b1e50e03885b0d2868c7390f91130d0a1cd95b4b75cdb1126",
"sampling_u": 0.4344415561103455,
"source_arrival_s": 488.6289999999999,
"source_request_id": "12588",
"source_row_index": 12588
},
{
"input_bin": 0,
"input_tokens": 736,
"prompt_sha256": "3514d193ff83b6d6589d52f24072b0159fa00d7b44ca8c6ec269515db4cfba89",
"sampling_u": 0.7257534252552246,
"source_arrival_s": 497.2449999999999,
"source_request_id": "12824",
"source_row_index": 12824
},
{
"input_bin": 0,
"input_tokens": 90,
"prompt_sha256": "915b3aee8d1288262c7f363f9035af19236e45540b748a9ced17c2d23f428f34",
"sampling_u": 0.7669962160527245,
"source_arrival_s": 520.1750000000002,
"source_request_id": "13350",
"source_row_index": 13350
},
{
"input_bin": 0,
"input_tokens": 697,
"prompt_sha256": "af93c55d3eb4f5161eb69410e73039e1d500a92190bff884e1b4040e491c8ed5",
"sampling_u": 0.594626944330893,
"source_arrival_s": 525.469,
"source_request_id": "13491",
"source_row_index": 13491
},
{
"input_bin": 1,
"input_tokens": 1470,
"prompt_sha256": "dd86d891c322f43310d4207c02f27a7d5e76885c420793164c76020f7ff166c4",
"sampling_u": 0.9612563305576178,
"source_arrival_s": 526.116,
"source_request_id": "13510",
"source_row_index": 13510
},
{
"input_bin": 1,
"input_tokens": 1592,
"prompt_sha256": "2ab08103a5ec0e01aea2f74f9215c52222ef8722a2fb9327bd9187a644df1c4b",
"sampling_u": 0.7977149220573079,
"source_arrival_s": 529.1809999999996,
"source_request_id": "13612",
"source_row_index": 13612
},
{
"input_bin": 5,
"input_tokens": 26776,
"prompt_sha256": "d09d69f9110f28c69d0767bdd501a8bdba167775ae1c6e6c53d15cf0fdcf5f6e",
"sampling_u": 0.9888861232028356,
"source_arrival_s": 546.9939999999997,
"source_request_id": "14026",
"source_row_index": 14026
},
{
"input_bin": 1,
"input_tokens": 1589,
"prompt_sha256": "973120a2be961f8f44523988a44f59b2071e82081a292cf37f10fb72d2874535",
"sampling_u": 0.8325378218001509,
"source_arrival_s": 549.6970000000001,
"source_request_id": "14094",
"source_row_index": 14094
},
{
"input_bin": 1,
"input_tokens": 1550,
"prompt_sha256": "8d899d80a6b9c47f414f6d65ed1638b5fc3bc782eaa19b2592878f35ff5484e7",
"sampling_u": 0.7285184471546545,
"source_arrival_s": 556.8310000000001,
"source_request_id": "14316",
"source_row_index": 14316
},
{
"input_bin": 2,
"input_tokens": 2501,
"prompt_sha256": "4a88c92c72a1fd9a653a7dbbd169eff30774f1b95cb5482f60c45532923d9f47",
"sampling_u": 0.38599257949920635,
"source_arrival_s": 558.3209999999999,
"source_request_id": "14359",
"source_row_index": 14359
},
{
"input_bin": 0,
"input_tokens": 741,
"prompt_sha256": "9c2341411206e2d28d6479ef20d5ce9e90daac574e72c4066130e2601731624c",
"sampling_u": 0.5930037066322071,
"source_arrival_s": 583.5829999999996,
"source_request_id": "15069",
"source_row_index": 15069
},
{
"input_bin": 3,
"input_tokens": 6044,
"prompt_sha256": "614a0d743b4cec35549b490110bc71d8fc3a22cd07c68de7281fc98270f81ee8",
"sampling_u": 0.4726761733329724,
"source_arrival_s": 594.8950000000004,
"source_request_id": "15377",
"source_row_index": 15377
}
],
"created_unix_s": 1784139549.1404486,
"load_contract": {
"binary_search": false,
"completion_tokens_override": 1,
"monotonicity_assumed": false,
"offered_rates": [
0.15,
0.2,
0.25,
0.3,
0.35,
0.4,
0.45,
0.5,
0.75,
1.0,
1.5
],
"only_mutated_variable": "uniform_arrival_timeline_scale",
"request_count_per_point": 64,
"target_pass_rate": 0.95
},
"primary_slo": "linear_8k_primary",
"rates": {
"r_0p150": {
"duration_s": 426.6666666666667,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.15,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p150.csv",
"request_count": 64,
"request_rate_recomputed": 0.15,
"sha256": "f35ce4424c17b2ee16144bb59011e467a07d3f57895716165fcdc73ba19839d9",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p200": {
"duration_s": 320.0,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.2,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p200.csv",
"request_count": 64,
"request_rate_recomputed": 0.2,
"sha256": "b81a48706a5c41222477aef8da357ae14c7b0bc19d35ea5ecc0c12b86d32590c",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p250": {
"duration_s": 256.0,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.25,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p250.csv",
"request_count": 64,
"request_rate_recomputed": 0.25,
"sha256": "721d622abb5d0a9080d6577063af910906b4e9928b402f48804ee55b25f108fc",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p300": {
"duration_s": 213.33333333333334,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.3,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p300.csv",
"request_count": 64,
"request_rate_recomputed": 0.3,
"sha256": "a5c8a7b93f314bf4365763fc0157597c582229993bb79b41b697ef3247495e49",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p350": {
"duration_s": 182.85714285714286,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.35,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p350.csv",
"request_count": 64,
"request_rate_recomputed": 0.35,
"sha256": "ec9debce28b76981f4f5363f0c77faa963c550a5d3ecbed327a304dc0d495242",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p400": {
"duration_s": 160.0,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.4,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p400.csv",
"request_count": 64,
"request_rate_recomputed": 0.4,
"sha256": "aea1cae38b09ddb0ad1a7d6da03bd6ad76afdaee3b2879b7cb65eba9b3ef5bab",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p450": {
"duration_s": 142.22222222222223,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.45,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p450.csv",
"request_count": 64,
"request_rate_recomputed": 0.44999999999999996,
"sha256": "672197deca834d9b9c5aea963448805a03e86495721d5eb270ccec2416faada2",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p500": {
"duration_s": 128.0,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.5,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p500.csv",
"request_count": 64,
"request_rate_recomputed": 0.5,
"sha256": "7cd2d16904e0fafb26a67718c4f2d7b6752934864219ab96813b8f62b47019d4",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_0p750": {
"duration_s": 85.33333333333333,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 0.75,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_0p750.csv",
"request_count": 64,
"request_rate_recomputed": 0.75,
"sha256": "e8b00f14b6c85ac8bd1c8fe3b83797f5c61903d70d8bbf6acfe9a1bd384db443",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_1p000": {
"duration_s": 64.0,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 1.0,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_1p000.csv",
"request_count": 64,
"request_rate_recomputed": 1.0,
"sha256": "7071e0247414af9214a1bf06f05a4865965568ba812bbae6713648e994a40a34",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
},
"r_1p500": {
"duration_s": 42.666666666666664,
"input_length_order_sha256": "d796fc955a9e7278512b56ce0544afd465d689e0060920c18839e85632360645",
"offered_request_rate": 1.5,
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/traces/r_1p500.csv",
"request_count": 64,
"request_rate_recomputed": 1.5,
"sha256": "86b03ef1fe87ad3c589f1585a279cd3136975009316e36e0910fe8340b972eee",
"source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5"
}
},
"schema": "qwen235b-prefill-fixed-cohort-protocol-v1",
"selection": {
"cohort_bin_quotas": [
27,
21,
4,
3,
5,
4
],
"cohort_size": 64,
"cohort_source_row_order_sha256": "d3945dc4a79840a0a157a73f10f2836006a3f0bbc5ecdbd657886865497affa5",
"input_bin_edges": [
0,
1024,
2048,
4096,
8192,
16384,
32769
],
"method": "length_stratified_smallest_sha256",
"seed": 2026071501,
"source_arrival_span_s": 585.9690000000005,
"source_bin_counts": [
6549,
5111,
916,
743,
1194,
896
],
"stable_order": [
"source_arrival_s",
"source_row_index"
]
},
"slo_variants": {
"legacy_step_1s_2s": {
"buckets": [
{
"max_input_tokens": 8191,
"threshold_ms": 1000.0
},
{
"threshold_ms": 2000.0
}
],
"description": "Original strict 1 s/2 s step SLO (null-capacity sensitivity)",
"kind": "step_ms"
},
"linear_10k": {
"description": "TTFT <= 1000 ms + input_tokens / 10000 tokens/s",
"intercept_ms": 1000.0,
"kind": "linear_ms",
"per_token_ms": 0.1
},
"linear_6k": {
"description": "TTFT <= 1000 ms + input_tokens / 6000 tokens/s",
"intercept_ms": 1000.0,
"kind": "linear_ms",
"per_token_ms": 0.16666666666666666
},
"linear_8k_primary": {
"description": "TTFT <= 1000 ms + input_tokens / 8000 tokens/s",
"intercept_ms": 1000.0,
"kind": "linear_ms",
"per_token_ms": 0.125
}
},
"source": {
"eligible_request_count": 15409,
"path": "/home/admin/cpfs/wjh/aituner/aituner/trace_windows/traces/thinking_w20260327_1000.jsonl",
"sha256": "f878e9af18f94dcfaced94a8e1e6b20a2f7d97d64aa862448025660dbbd965b2"
}
}

View File

@@ -0,0 +1,204 @@
# Qwen3-235B FP8Frontier 与 community vLLM config rank 对比
## 结论
在本实验限定的 Qwen3-235B-A22B-FP8、H20、prefill-only workload 上,经过 best-effort 补齐 FP8/MoE serving path 后Frontier **能够选对满足 SLO 的 per-GPU throughput 最优配置**`TP=4, MBT=16384``MNS=64/128` 并列。细化网格上的 Spearman rank correlation 为 `0.9487`20 个可比较非 tie config pair 全部同序,选择 regret 为 `0`
但这不等于 Frontier 已精确模拟系统。原始粗网格给出 56/56 load labels 正确、Spearman=1 的表象;加入 decision-boundary rates 后Frontier 在 33 个被测试边界点中产生 6 个 false-infeasible均为两个 MNS 配置上可复现的稳定误差。它系统性地把三个 action family 的最大经验 SLO throughput 低估一个 `0.05 req/s` 档位:
| TP | MBT | Frontier capacity | real capacity | real v2→v3 transition | 结论 |
|---:|---:|---:|---:|---|---|
| 4 | 8192 | 0.15 | 0.20 | `0.20: 61→61`, `0.25: 60→60` | 稳定低估 0.05 |
| 4 | 16384 | 0.25 | 0.30 | `0.30: 61→61`, `0.35: 59→59` | 稳定低估 0.05 |
| 8 | 8192 | 0.40 | 0.45 | `0.45: 61→61`, `0.50: 60→60` | 稳定低估 0.05 |
| 8 | 16384 | 0.45 | 0.45 | v2: `0.45: 61`, `0.50: 60` | 边界正确 |
表中 pass count 分母均为 64经验可行门槛为 61。每一行的两个 MNS 结果一致v3 对前三行的两个 MNS 都做了 fresh-server 复测。
因此,对这个 case 最准确的判断是:
- Frontier 有较好的 **decision/ordinal fidelity**,不能再用“它找不到最优 config”作为该 case 的研究 premise。
- Frontier 没有达到 **capacity calibration fidelity**,也没有完全学对 action effect它适合筛选但不能在 SLO knee 附近代替真机 capacity certification。
- 粗网格上的“完全匹配”并不足以证明 simulator solve 了问题;它会同时掩盖绝对容量误差和错误的 action differential。
## 固定实验口径
| 项目 | 固定值 |
|---|---|
| model | `models/Qwen/Qwen3-235B-A22B-FP8` |
| real runtime | community vLLM 0.10.2 |
| hardware | dash08×NVIDIA H20 |
| workload | prefill-onlyraw completionoutput tokens=1prefix cache disabled |
| config surface | `TP∈{4,8} × MNS∈{64,128} × MBT∈{8192,16384}` |
| cohort | 64 个固定 prompt所有 config/rate 的 request ID、顺序、输入长度完全相同 |
| input length | p50=1478p95=18439max=29385mean=3511.55 tokens |
| rates | 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.75, 1.00, 1.50 req/s |
| primary SLO | `TTFT ≤ 1000 ms + input_tokens / 8000 tokens/s`,至少 95% requests 通过 |
| capacity | 所有 64 requests 完整结束后,最大被测试可行 offered rate不做 early stop/二分单调假设 |
| primary objective | capacity / TP即 per-GPU request throughput |
cohort 只通过 uniform arrival scaling 改变 load不随 rate 重新抽样。该设计消除了 workload composition confound但 64-request 固定 cohort 的结果仍是经验容量,不代表总体 workload distribution 的置信界。
## Frontier best-effort 公平性
这里比较的不是未经修改的 stock Frontier。为了避免“故意把 baseline 调差”的质疑,我们在真机 ranking 冻结前完成并冻结了以下补齐:
- 使用 community vLLM 0.10.2 在 H20 上得到的真实 FP8 operator profilesquantization contract 为 FP8 W8A8 dynamic activation、BF16 output/accumulation而不是用 BF16 compute profile 冒充 FP8。
- 补齐 Qwen MoE community serving path、MoE plan 与 serving entrypoint。
- 按真实 expert route 构造 EP local lanes并以最慢 lane 作为同步 critical path。
- 修正 TP/EP-aware profile cache key避免不同 topology 错用缓存结果。
- 使用 batched lane prediction与逐 lane exact path 做 byte-identical equivalence 后才用于完整 sweep模拟运行加速约 2.57×。
- KV block capacity 使用真机测量值TP4=26101TP8=62351。
- Frontier refined 8×11=88 个结果在真机运行前冻结;旧 rates 的 56 个结果与先前 freeze byte-identical。没有用真机 v2/v3 结果回调 simulator 参数。
相关实现与审计入口包括:
- `fixed_cohort_rank.py`
- `frontier_fp8_profile_contract.patch`
- `frontier_moe_serving_path.patch`
- `frontier_moe_ep_prefill_exact_cache.patch`
- `frontier_moe_ep_prefill_use_batched_lanes.patch`
- `frontier_moe_ep_prefill_batched_lane_predict.patch`
这说明结果反映的是“给 Frontier 当前抽象最大的合理帮助以后,它能做到什么”,不是 stock baseline 的缺功能测试。
## 三阶段实验结果
### v1blind 全表面、粗网格
Frontier 先冻结,再运行 community vLLM 的 8 configs × 7 rates
- 56/56 config-load primary feasibility labels 正确。
- raw capacity 与 per-GPU efficiency 的 top set 均完全匹配。
- Spearman=120/20 非 tie pair 同序top-1 regret=0。
- 但 3584 个 request 的 TTFT MAE 为 124.20 msMAPE 为 20.98%,平均 `predicted-real=+93.20 ms`
- request-level SLO label accuracy 为 96.82%Frontier/real failure-set Jaccard 仅 0.7569103 false fail、11 false pass
因此aggregate capacity 完全匹配时,具体失败请求和 request latency mechanism 已明显不匹配。
### v2反向 config/rate 顺序,加入细化边界
v2 只重跑预注册 SLO transition 及中间 refinement rates共 8 configs、33 loads、2112 requests
- 0 server/request failure0 early stop0 monotonicity violation所有 trace hash 匹配 refined protocol。
- per-GPU top set 仍精确匹配:`TP4, MBT16384, MNS64/128`
- Frontier/real per-GPU capacity Spearman=0.948720/20 非 tie pair 同序regret=0。
- 33 个边界点中 27 个 feasibility label 正确6 个错误全部是 false-infeasible没有 false-feasible。
- pass-rate MAE=1.278 percentage points平均 `Frontier-real=-1.278 pp`
- 2112 个 request 的 TTFT MAE=83.15 msMAPE=16.34%,平均 `predicted-real=+53.02 ms`
- request SLO label accuracy=98.34%failure-set Jaccard=0.7535。
v1 与 v2 的 19 个共同 config-load 点包含 1216 个相同 requests
- 真机 trial-to-trial TTFT MAE=14.05 msabsolute-error p95=57.94 ms平均 v2-v1=+1.75 ms。
- primary request label 只有 1/1216 flip19/19 load labels 和 8/8 原始网格 capacity 全部稳定。
- Frontier 有 23 个 request label 在两次真机 trial 中都稳定预测错误。
这排除了 config order 和普通 runtime noise 是主要 residual 来源。
### v3只复测 v2 的全部争议 transition
v3 的选择规则是 v2 中所有 Frontier/real primary capacity 不同的 config它是 data-dependent replication不作为新的 blind primary result。共 6 configs、12 loads、768 requests
- 0 failure/early-stop/monotonicity violationtrace hash 全部匹配。
- v2 与 v3 的 768/768 primary request labels 完全一致failure-set Jaccard=1.0。
- 12/12 config-load labels、6/6 capacity boundaries 完全一致。
- 真机 trial-to-trial TTFT MAE=14.98 msabsolute-error p95=62.52 ms平均 v3-v2=+1.55 ms。
- Frontier 在该争议集合上仍有 6/12 false-infeasible18 个 request labels 在 v2/v3 都稳定预测错误。
这证明细化边界误差不是一次随机跨阈值。
## 系统层面的理解
### 1. 绝对误差与 action differential 必须分开
如果只是一个全局 latency scale 偏差,所有 topology 下 MBT 的相对收益应基本保持。数据并非如此:
- TP4Frontier 预测 MBT `8192→16384` 的 capacity 增益为 `0.15→0.25`,即 +0.10 req/s真机为 `0.20→0.30`,同样 +0.10。
- TP8Frontier 预测 `0.40→0.45`,即 +0.05;真机为 `0.45→0.45`,实际增益为 0。
也就是说Frontier 在 TP4 学对了 MBT differential、错了 absolute offset在 TP8 则连 differential 都错了。该误差依赖 TP/EP topology不能靠一个全局 calibration factor 修复。这正是 execution configuration 与 scheduler state/action effect 相互耦合的证据。
### 2. 决策正确不代表 simulator 对系统有正确解释
per-GPU objective 下TP4/MBT16384 与其他配置有足够大的 margin所以即便 capacity 每类偏差 0.05,最优 config 仍不变。raw single-replica objective 下Frontier 只把 TP8/MBT16384 列为 top set而真机四个 TP8 config 全部在 0.45 并列;虽然从 Frontier top set 中任选仍是 zero regret但它虚构了 TP8 上 MBT 的收益。
因此需要同时报告:
1. selected-config regret
2. rank/order fidelity
3. capacity error
4. action differential error
5. request/failure-set fidelity。
只报其中任意一层都会得到偏颇结论。
### 3. residual 是状态/长度相关尾部误差,不是 profile 全局比例错误
v2 中 Frontier 的平均 signed TTFT error 随输入长度显著变化:
| input tokens | mean `predicted-real` | MAE |
|---|---:|---:|
| [0,1024) | -3.61 ms | 40.74 ms |
| [1024,2048) | +44.45 ms | 57.63 ms |
| [2048,4096) | +91.21 ms | 109.73 ms |
| [4096,8192) | +138.80 ms | 140.80 ms |
| [8192,16384) | +237.37 ms | 237.37 ms |
| [16384,32769) | +147.36 ms | 240.89 ms |
短请求甚至略偏乐观,中长请求明显偏保守;与此同时 aggregate load label 只由 12 个请求是否跨 SLO 阈值决定。这更像 batch composition、critical-lane/EP synchronization 和 queueing state 共同造成的 residual而不是 operator profile 全局慢了固定百分比。
### 4. MNS 在该 workload 上是非活跃维度,但不能外推
在所有共同测试点中MNS=64/128 的 pass counts 基本逐点相同Frontier 也预测 tie。对这个 prefill-only、固定 cohort、当前 concurrency 范围MBT/token budget 和 topology 是活跃因素MNS 没有形成独立瓶颈。但 decode/mixed workload 的 sequence residency、KV pressure 和 iteration cadence 不同,不能据此宣称 MNS 普遍无效。
## 对研究方向的判断
这个 case 不支持以下论文 premise
> Frontier 类 simulator 无法为 Qwen3-235B FP8 选择最优配置,所以必须由一个新 tuner 替代。
在公平补齐 FP8/MoE path 后,它已经能以 zero regret 选出该 workload 的 per-GPU 最优配置。继续用这一 case 证明“simulator 排序失败”会被 reviewer 用本组数据直接反驳。
本 case 真正揭示、且更有 systems 内容的问题是:
> 一个 simulator 可以在 coarse grid 上拥有完美 rank并在最终 objective 上 zero regret同时仍错误表达 topology-dependent action differential 和 SLO knee我们如何判断一个模拟决策何时足够可信何时必须回到真机
可继续发展的研究问题包括:
- decision-aware uncertainty不追求每条 TTFT 精确,而是给 config ordering/capacity boundary 提供可校验置信区间。
- counterexample-guided measurement只在模拟 action margin 小或 differential 不稳定的 boundary/state 上调用真机,而不是全表面 sweep。
- topology-aware residual model显式建模 TP/EP、batch composition、critical lane 与输入长度共同造成的尾部偏差,避免全局比例校准。
- distinguish safe screening from certificationFrontier 当前在该 case 是保守筛选器0 false-feasible但不是精确 capacity certifier。
下一步要判断这是否足以构成更一般的 research premise应转向已计划的 Qwen3-235B decode-only 和 mixed prefill/decode case。它们会激活 KV state、replica residency、iteration cadence 和 communication overlap如果只有这些 case 出现稳定排序/regret 错误,研究问题就应围绕 simulator 在 stateful execution/scheduling coupling 下的决策失效,而不是泛化地宣称 simulator 不工作。
## 证据与复现
本目录保存紧凑 comparison、trial-stability 与 manifest完整 request-level freeze 保留在 dash0
`/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1`
关键 artifact SHA256
| artifact | SHA256 |
|---|---|
| refined protocol manifest | `695c0500b9a5d51cdd4e39b722e3c9db15f8bed48795454f4d0d6442d24929cb` |
| Frontier refined freeze | `318b3c7742c8371191e2dbd429f0e0e63c1a42f4806a599e018b0548cb9d6336` |
| v1 blind comparison | `fc87c54179b891e2148a738dbdfeed328595fa32198f2248c892c890625c9fe2` |
| v2 run manifest | `cf1a722043f42ff8f435e3521d0b2b2892a8ea0922e1e4870ec10fba26c0f998` |
| v2 community freeze | `5a9c59fc4082db47bec7b59432e9ebe9f907f70baa95a9f0d3480e85f8992ff2` |
| v2 comparison | `d21f2dd846c4d8361668f312f0c063f2e8f949ec7a9d3b19100c813b3da3b586` |
| v1-v2 stability | `a522189618ad26df68f18d2d9b67c38b1fe59c5c8f13af633eced0b092f8e06d` |
| v3 run manifest | `0f8908671b8903e516f0dc8e68f556483a15ff7ec51f7cfbc18a9c74b3a5ba61` |
| v3 community freeze | `4edd25edde3d7508d8feb9b024b5bca4ba6a5f25335ac84008f09c524ed7bc34` |
| v3 comparison | `e04f40ead995e007d64b2f6f0f2731999204907770f05b4c2d037d224e49c573` |
| v2-v3 stability | `7be2ff27dd0393c99f0512bd6a0b7b35eb404ee0bfd22bd3ec33a4b8ce14748b` |
## 限制
- 只测试了一个模型、H20、community vLLM 0.10.2 和 prefill-only workload。
- 固定 64-request cohort 适合 simulator-vs-real paired comparison但不足以给总体 workload distribution 的 95% SLO capacity 做统计认证。
- v3 是在观察 v2 后选择全部争议点的 replication不是 blind discovery dataset。
- operator profiles、KV capacity 和 MoE route contract 本身使用了真机信息;本结果评估的是 best-effort simulator 的决策能力与剩余误差,不是 zero-measurement transfer。
- 没有测试 decode、mixed phase、prefix caching、multi-replica placement、异构 GPU 或跨节点通信。

View File

@@ -0,0 +1,583 @@
{
"configs": [
{
"config": {
"expert_parallel": true,
"mbt": 16384,
"mns": 128,
"name": "tp8_mns128_mbt16384",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp8_mns128_mbt16384/engine.log",
"execution_index": 1,
"rate_order": [
0.75,
0.5,
0.35,
0.45,
0.4
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp8_mns128_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": true,
"mbt": 16384,
"mns": 64,
"name": "tp8_mns64_mbt16384",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp8_mns64_mbt16384/engine.log",
"execution_index": 2,
"rate_order": [
0.5,
0.35,
0.45,
0.4
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp8_mns64_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 16384,
"mns": 64,
"name": "tp4_mns64_mbt16384",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp4_mns64_mbt16384/engine.log",
"execution_index": 3,
"rate_order": [
0.35,
0.25,
0.3
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp4_mns64_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": true,
"mbt": 8192,
"mns": 64,
"name": "tp8_mns64_mbt8192",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp8_mns64_mbt8192/engine.log",
"execution_index": 4,
"rate_order": [
0.5,
0.35,
0.45,
0.4
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp8_mns64_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": true,
"mbt": 8192,
"mns": 128,
"name": "tp8_mns128_mbt8192",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp8_mns128_mbt8192/engine.log",
"execution_index": 5,
"rate_order": [
0.5,
0.35,
0.45,
0.4
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp8_mns128_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 8192,
"mns": 128,
"name": "tp4_mns128_mbt8192",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp4_mns128_mbt8192/engine.log",
"execution_index": 6,
"rate_order": [
0.15,
0.35,
0.25,
0.3,
0.2
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp4_mns128_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 8192,
"mns": 64,
"name": "tp4_mns64_mbt8192",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp4_mns64_mbt8192/engine.log",
"execution_index": 7,
"rate_order": [
0.25,
0.35,
0.15,
0.3,
0.2
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp4_mns64_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 16384,
"mns": 128,
"name": "tp4_mns128_mbt16384",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/runs/tp4_mns128_mbt16384/engine.log",
"execution_index": 8,
"rate_order": [
0.35,
0.25,
0.3
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/results/tp4_mns128_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
}
],
"created_unix_s": 1784142580.9845052,
"execution_contract": {
"config_order": "reverse_of_first_pass",
"estimated_replay_wall_seconds_excluding_model_loads": 6639.746031746032,
"fallback_when_no_transition": "top_two_if_all_feasible_else_bottom_two",
"one_fresh_server_launch_per_config": true,
"rate_order": "reverse_first_pass_boundary_order_then_descending_refinement_rates",
"schema": "community-qwen235b-boundary-repeat-selection-v4",
"selection": "all_adjacent_pre_registered_slo_feasibility_transitions_plus_between_rate_refinement",
"selections": [
{
"added_refinement_rates": [
0.4,
0.45
],
"boundary_rates": [
0.35,
0.5,
0.75
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [
0.35,
0.5
],
"linear_6k": [
0.5,
0.75
],
"linear_8k_primary": [
0.35,
0.5
]
},
"config": "tp8_mns128_mbt16384",
"excluded_unrankable_slos": [
"legacy_step_1s_2s"
],
"repeat_order": [
0.75,
0.5,
0.35,
0.45,
0.4
],
"selected_rates": [
0.35,
0.4,
0.45,
0.5,
0.75
],
"source_relative_order": [
0.35,
0.5,
0.75
]
},
{
"added_refinement_rates": [
0.4,
0.45
],
"boundary_rates": [
0.35,
0.5
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [
0.35,
0.5
],
"linear_6k": [
0.35,
0.5
],
"linear_8k_primary": [
0.35,
0.5
]
},
"config": "tp8_mns64_mbt16384",
"excluded_unrankable_slos": [
"legacy_step_1s_2s"
],
"repeat_order": [
0.5,
0.35,
0.45,
0.4
],
"selected_rates": [
0.35,
0.4,
0.45,
0.5
],
"source_relative_order": [
0.35,
0.5
]
},
{
"added_refinement_rates": [
0.3
],
"boundary_rates": [
0.25,
0.35
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [],
"linear_6k": [
0.25,
0.35
],
"linear_8k_primary": [
0.25,
0.35
]
},
"config": "tp4_mns64_mbt16384",
"excluded_unrankable_slos": [
"linear_10k",
"legacy_step_1s_2s"
],
"repeat_order": [
0.35,
0.25,
0.3
],
"selected_rates": [
0.25,
0.3,
0.35
],
"source_relative_order": [
0.25,
0.35
]
},
{
"added_refinement_rates": [
0.4,
0.45
],
"boundary_rates": [
0.35,
0.5
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [
0.35,
0.5
],
"linear_6k": [
0.35,
0.5
],
"linear_8k_primary": [
0.35,
0.5
]
},
"config": "tp8_mns64_mbt8192",
"excluded_unrankable_slos": [
"legacy_step_1s_2s"
],
"repeat_order": [
0.5,
0.35,
0.45,
0.4
],
"selected_rates": [
0.35,
0.4,
0.45,
0.5
],
"source_relative_order": [
0.35,
0.5
]
},
{
"added_refinement_rates": [
0.4,
0.45
],
"boundary_rates": [
0.35,
0.5
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [
0.35,
0.5
],
"linear_6k": [
0.35,
0.5
],
"linear_8k_primary": [
0.35,
0.5
]
},
"config": "tp8_mns128_mbt8192",
"excluded_unrankable_slos": [
"legacy_step_1s_2s"
],
"repeat_order": [
0.5,
0.35,
0.45,
0.4
],
"selected_rates": [
0.35,
0.4,
0.45,
0.5
],
"source_relative_order": [
0.35,
0.5
]
},
{
"added_refinement_rates": [
0.2,
0.3
],
"boundary_rates": [
0.15,
0.25,
0.35
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [],
"linear_6k": [
0.25,
0.35
],
"linear_8k_primary": [
0.15,
0.25
]
},
"config": "tp4_mns128_mbt8192",
"excluded_unrankable_slos": [
"linear_10k",
"legacy_step_1s_2s"
],
"repeat_order": [
0.15,
0.35,
0.25,
0.3,
0.2
],
"selected_rates": [
0.15,
0.2,
0.25,
0.3,
0.35
],
"source_relative_order": [
0.25,
0.35,
0.15
]
},
{
"added_refinement_rates": [
0.2,
0.3
],
"boundary_rates": [
0.15,
0.25,
0.35
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [],
"linear_6k": [
0.25,
0.35
],
"linear_8k_primary": [
0.15,
0.25
]
},
"config": "tp4_mns64_mbt8192",
"excluded_unrankable_slos": [
"linear_10k",
"legacy_step_1s_2s"
],
"repeat_order": [
0.25,
0.35,
0.15,
0.3,
0.2
],
"selected_rates": [
0.15,
0.2,
0.25,
0.3,
0.35
],
"source_relative_order": [
0.15,
0.35,
0.25
]
},
{
"added_refinement_rates": [
0.3
],
"boundary_rates": [
0.25,
0.35
],
"boundary_rates_by_slo": {
"legacy_step_1s_2s": [],
"linear_10k": [],
"linear_6k": [
0.25,
0.35
],
"linear_8k_primary": [
0.25,
0.35
]
},
"config": "tp4_mns128_mbt16384",
"excluded_unrankable_slos": [
"linear_10k",
"legacy_step_1s_2s"
],
"repeat_order": [
0.35,
0.25,
0.3
],
"selected_rates": [
0.25,
0.3,
0.35
],
"source_relative_order": [
0.25,
0.35
]
}
],
"unrankable_slo_policy": "exclude_per_config_surfaces_with_no_feasible_load",
"warmup": {
"latencies_discarded": true,
"method": "one_disjoint_request_per_input_length_bin",
"request_count": 6,
"seed": 2026071504
}
},
"frontier_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/frontier-v4-refined-rates-batched-par4/frontier_ranking_frozen.json",
"sha256": "318b3c7742c8371191e2dbd429f0e0e63c1a42f4806a599e018b0548cb9d6336"
},
"model": {
"config_sha256": "702c46d431bb984db9035a1225186bbfdb52c0d19c82104df4a37cd005e0369e",
"path": "/home/admin/cpfs/wjh/models/Qwen/Qwen3-235B-A22B-FP8"
},
"protocol_manifest": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/protocol_manifest.json",
"sha256": "695c0500b9a5d51cdd4e39b722e3c9db15f8bed48795454f4d0d6442d24929cb"
},
"repeat_of": {
"first_pass_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/community_ranking_frozen.json",
"sha256": "055fe2cf1e0c01a0503145a272097b8ceda257f1b86bdae39abf17277d9a18d9"
},
"source_manifest": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/run_manifest.json",
"sha256": "6da81a34140d801cb6055e089bed7a12adb14c09dc87ca6dfa7f79df4ac34920"
}
},
"repository": {
"commit": "501ceb817122747a23dc6d9fdc9d71dcb3aa8e72",
"status_porcelain": []
},
"runtime": {
"python": "/usr/bin/python3.12",
"vllm": "/tmp/wjh-frontier-vllm0102-smoke/.venv/bin/vllm"
},
"schema": "community-qwen235b-prefill-fixed-cohort-v1",
"studies": {
"4": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json",
"sha256": "ea0d7d31e4a5102cf530e900754debb1d8b18f860748ead88ebf7de57d92faa3"
},
"8": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json",
"sha256": "48c4d917b5531f742a05bab3bcd5c880f2eaddf9992b7629e9f48d8e00fd1906"
}
}
}

View File

@@ -0,0 +1,732 @@
{
"created_unix_s": 1784153851.3812342,
"inputs": {
"first_community_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/community_ranking_frozen.json",
"sha256": "5a9c59fc4082db47bec7b59432e9ebe9f907f70baa95a9f0d3480e85f8992ff2"
},
"frontier_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/frontier-v4-refined-rates-batched-par4/frontier_ranking_frozen.json",
"sha256": "318b3c7742c8371191e2dbd429f0e0e63c1a42f4806a599e018b0548cb9d6336"
},
"repeat_community_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/community_ranking_frozen.json",
"sha256": "4edd25edde3d7508d8feb9b024b5bca4ba6a5f25335ac84008f09c524ed7bc34"
}
},
"primary_original_grid_capacity_stability": {
"per_gpu_efficiency": {
"capacity_field": "maximum_tested_feasible_request_rate_per_gpu",
"frontier_capacity": {
"tp4_mns128_mbt16384": 0.075,
"tp4_mns128_mbt8192": 0.05,
"tp4_mns64_mbt16384": 0.075,
"tp4_mns64_mbt8192": 0.05,
"tp8_mns128_mbt16384": 0.05625,
"tp8_mns128_mbt8192": 0.05625,
"tp8_mns64_mbt16384": 0.05625,
"tp8_mns64_mbt8192": 0.05625
},
"frontier_top1_set": [
"tp4_mns128_mbt16384",
"tp4_mns64_mbt16384"
],
"pairwise_ordering": {
"accuracy": 1.0,
"comparable_non_tied_pairs": 12,
"correct_pairs": 12,
"frontier_tied_pairs": 3,
"real_tied_pairs": 3
},
"rankable": true,
"real_capacity": {
"tp4_mns128_mbt16384": 0.075,
"tp4_mns128_mbt8192": 0.05,
"tp4_mns64_mbt16384": 0.075,
"tp4_mns64_mbt8192": 0.05,
"tp8_mns128_mbt8192": 0.05625,
"tp8_mns64_mbt8192": 0.05625
},
"real_top1_set": [
"tp4_mns128_mbt16384",
"tp4_mns64_mbt16384"
],
"spearman_rank_correlation": 1.0,
"top1_regret_fraction_best_tie_break": 0.0,
"top1_regret_fraction_worst_tie_break": 0.0,
"top1_set_intersection": [
"tp4_mns128_mbt16384",
"tp4_mns64_mbt16384"
],
"top1_set_match": true
},
"records": [
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp4_mns128_mbt16384",
"first_capacity": 0.3,
"first_next_tested_rate": 0.35,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
},
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp4_mns128_mbt8192",
"first_capacity": 0.2,
"first_next_tested_rate": 0.25,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
},
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp4_mns64_mbt16384",
"first_capacity": 0.3,
"first_next_tested_rate": 0.35,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
},
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp4_mns64_mbt8192",
"first_capacity": 0.2,
"first_next_tested_rate": 0.25,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
},
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp8_mns128_mbt8192",
"first_capacity": 0.45,
"first_next_tested_rate": 0.5,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
},
{
"boundary_repeated": true,
"capacity_still_feasible": true,
"config": "tp8_mns64_mbt8192",
"first_capacity": 0.45,
"first_next_tested_rate": 0.5,
"next_rate_still_infeasible": true,
"stable_original_grid_capacity": true
}
],
"single_replica_raw_capacity": {
"capacity_field": "maximum_tested_feasible_request_rate",
"frontier_capacity": {
"tp4_mns128_mbt16384": 0.3,
"tp4_mns128_mbt8192": 0.2,
"tp4_mns64_mbt16384": 0.3,
"tp4_mns64_mbt8192": 0.2,
"tp8_mns128_mbt16384": 0.45,
"tp8_mns128_mbt8192": 0.45,
"tp8_mns64_mbt16384": 0.45,
"tp8_mns64_mbt8192": 0.45
},
"frontier_top1_set": [
"tp8_mns128_mbt16384",
"tp8_mns128_mbt8192",
"tp8_mns64_mbt16384",
"tp8_mns64_mbt8192"
],
"pairwise_ordering": {
"accuracy": 1.0,
"comparable_non_tied_pairs": 12,
"correct_pairs": 12,
"frontier_tied_pairs": 3,
"real_tied_pairs": 3
},
"rankable": true,
"real_capacity": {
"tp4_mns128_mbt16384": 0.3,
"tp4_mns128_mbt8192": 0.2,
"tp4_mns64_mbt16384": 0.3,
"tp4_mns64_mbt8192": 0.2,
"tp8_mns128_mbt8192": 0.45,
"tp8_mns64_mbt8192": 0.45
},
"real_top1_set": [
"tp8_mns128_mbt8192",
"tp8_mns64_mbt8192"
],
"spearman_rank_correlation": 0.9933992677987828,
"top1_regret_fraction_best_tie_break": 0.0,
"top1_regret_fraction_worst_tie_break": 0.0,
"top1_set_intersection": [
"tp8_mns128_mbt8192",
"tp8_mns64_mbt8192"
],
"top1_set_match": false
},
"stable_configs": 6,
"total_configs": 6
},
"real_trial_ttft_ms": {
"absolute_error_p95": 62.524845998268574,
"mae": 14.980625974885697,
"paired_successful_requests": 768,
"repeat_minus_first_mean": 1.5463623839574818
},
"schema": "community-qwen235b-trial-stability-v1",
"shared_config_load_points": 12,
"shared_requests": 768,
"variants": {
"legacy_step_1s_2s": {
"config_load_records": [
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 56,
"offered_request_rate": 0.3,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 54,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 56
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.2,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 57,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 57
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 56,
"offered_request_rate": 0.3,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 54,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 56
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.2,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 57,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 57
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 57,
"frontier_feasible": false,
"frontier_passed_request_count": 54,
"offered_request_rate": 0.45,
"repeat_feasible": false,
"repeat_passed_request_count": 57
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 53,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 56
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 57,
"frontier_feasible": false,
"frontier_passed_request_count": 54,
"offered_request_rate": 0.45,
"repeat_feasible": false,
"repeat_passed_request_count": 57
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 53,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 56
}
],
"frontier_correct_first": 742,
"frontier_correct_repeat": 742,
"frontier_load_label_accuracy_first": 1.0,
"frontier_load_label_accuracy_repeat": 1.0,
"frontier_request_label_accuracy_first": 0.9661458333333334,
"frontier_request_label_accuracy_repeat": 0.9661458333333334,
"frontier_wrong_both_stable_real_label": 26,
"matching_real_trial_labels": 768,
"real_trial_failure_set_jaccard": 1.0,
"real_trial_label_flips": 0,
"real_trial_load_label_accuracy": 1.0,
"real_trial_request_label_accuracy": 1.0,
"shared_config_load_points": 12,
"shared_requests": 768
},
"linear_10k": {
"config_load_records": [
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 56,
"offered_request_rate": 0.3,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 55,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 56
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.2,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 56,
"offered_request_rate": 0.3,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 56,
"frontier_feasible": false,
"frontier_passed_request_count": 55,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 56
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.2,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 58,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 58
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 58,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
}
],
"frontier_correct_first": 726,
"frontier_correct_repeat": 726,
"frontier_load_label_accuracy_first": 0.8333333333333334,
"frontier_load_label_accuracy_repeat": 0.8333333333333334,
"frontier_request_label_accuracy_first": 0.9453125,
"frontier_request_label_accuracy_repeat": 0.9453125,
"frontier_wrong_both_stable_real_label": 42,
"matching_real_trial_labels": 768,
"real_trial_failure_set_jaccard": 1.0,
"real_trial_label_flips": 0,
"real_trial_load_label_accuracy": 1.0,
"real_trial_request_label_accuracy": 1.0,
"shared_config_load_points": 12,
"shared_requests": 768
},
"linear_6k": {
"config_load_records": [
{
"config": "tp4_mns128_mbt16384",
"first_feasible": true,
"first_passed_request_count": 62,
"frontier_feasible": true,
"frontier_passed_request_count": 61,
"offered_request_rate": 0.3,
"repeat_feasible": true,
"repeat_passed_request_count": 62
},
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 62,
"frontier_feasible": true,
"frontier_passed_request_count": 62,
"offered_request_rate": 0.2,
"repeat_feasible": true,
"repeat_passed_request_count": 62
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": true,
"frontier_passed_request_count": 62,
"offered_request_rate": 0.25,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": true,
"first_passed_request_count": 62,
"frontier_feasible": true,
"frontier_passed_request_count": 61,
"offered_request_rate": 0.3,
"repeat_feasible": true,
"repeat_passed_request_count": 62
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 62,
"frontier_feasible": true,
"frontier_passed_request_count": 62,
"offered_request_rate": 0.2,
"repeat_feasible": true,
"repeat_passed_request_count": 62
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": true,
"frontier_passed_request_count": 62,
"offered_request_rate": 0.25,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
}
],
"frontier_correct_first": 760,
"frontier_correct_repeat": 760,
"frontier_load_label_accuracy_first": 0.8333333333333334,
"frontier_load_label_accuracy_repeat": 0.8333333333333334,
"frontier_request_label_accuracy_first": 0.9895833333333334,
"frontier_request_label_accuracy_repeat": 0.9895833333333334,
"frontier_wrong_both_stable_real_label": 8,
"matching_real_trial_labels": 768,
"real_trial_failure_set_jaccard": 1.0,
"real_trial_label_flips": 0,
"real_trial_load_label_accuracy": 1.0,
"real_trial_request_label_accuracy": 1.0,
"shared_config_load_points": 12,
"shared_requests": 768
},
"linear_8k_primary": {
"config_load_records": [
{
"config": "tp4_mns128_mbt16384",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.3,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp4_mns128_mbt16384",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.2,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp4_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.3,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp4_mns64_mbt16384",
"first_feasible": false,
"first_passed_request_count": 59,
"frontier_feasible": false,
"frontier_passed_request_count": 57,
"offered_request_rate": 0.35,
"repeat_feasible": false,
"repeat_passed_request_count": 59
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.2,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp4_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.25,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns128_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": true,
"first_passed_request_count": 61,
"frontier_feasible": false,
"frontier_passed_request_count": 60,
"offered_request_rate": 0.45,
"repeat_feasible": true,
"repeat_passed_request_count": 61
},
{
"config": "tp8_mns64_mbt8192",
"first_feasible": false,
"first_passed_request_count": 60,
"frontier_feasible": false,
"frontier_passed_request_count": 59,
"offered_request_rate": 0.5,
"repeat_feasible": false,
"repeat_passed_request_count": 60
}
],
"frontier_correct_first": 750,
"frontier_correct_repeat": 750,
"frontier_load_label_accuracy_first": 0.5,
"frontier_load_label_accuracy_repeat": 0.5,
"frontier_request_label_accuracy_first": 0.9765625,
"frontier_request_label_accuracy_repeat": 0.9765625,
"frontier_wrong_both_stable_real_label": 18,
"matching_real_trial_labels": 768,
"real_trial_failure_set_jaccard": 1.0,
"real_trial_label_flips": 0,
"real_trial_load_label_accuracy": 1.0,
"real_trial_request_label_accuracy": 1.0,
"shared_config_load_points": 12,
"shared_requests": 768
}
}
}

View File

@@ -0,0 +1,345 @@
{
"configs": [
{
"config": {
"expert_parallel": false,
"mbt": 16384,
"mns": 128,
"name": "tp4_mns128_mbt16384",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp4_mns128_mbt16384/engine.log",
"execution_index": 1,
"rate_order": [
0.3,
0.35
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp4_mns128_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 8192,
"mns": 64,
"name": "tp4_mns64_mbt8192",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp4_mns64_mbt8192/engine.log",
"execution_index": 2,
"rate_order": [
0.2,
0.25
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp4_mns64_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 8192,
"mns": 128,
"name": "tp4_mns128_mbt8192",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp4_mns128_mbt8192/engine.log",
"execution_index": 3,
"rate_order": [
0.2,
0.25
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp4_mns128_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
},
{
"config": {
"expert_parallel": true,
"mbt": 8192,
"mns": 128,
"name": "tp8_mns128_mbt8192",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp8_mns128_mbt8192/engine.log",
"execution_index": 4,
"rate_order": [
0.45,
0.5
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp8_mns128_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": true,
"mbt": 8192,
"mns": 64,
"name": "tp8_mns64_mbt8192",
"num_gpu_blocks": 62351,
"tp": 8
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp8_mns64_mbt8192/engine.log",
"execution_index": 5,
"rate_order": [
0.45,
0.5
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp8_mns64_mbt8192.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json"
},
{
"config": {
"expert_parallel": false,
"mbt": 16384,
"mns": 64,
"name": "tp4_mns64_mbt16384",
"num_gpu_blocks": 26101,
"tp": 4
},
"engine_log_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/runs/tp4_mns64_mbt16384/engine.log",
"execution_index": 6,
"rate_order": [
0.3,
0.35
],
"result_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v3-primary-disputed/results/tp4_mns64_mbt16384.json",
"study_path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json"
}
],
"created_unix_s": 1784150504.0091515,
"execution_contract": {
"config_order": "reverse_of_first_pass",
"estimated_replay_wall_seconds_excluding_model_loads": 2484.8253968253966,
"fallback_when_no_transition": "top_two_if_all_feasible_else_bottom_two",
"one_fresh_server_launch_per_config": true,
"rate_order": "reverse_first_pass_boundary_order_then_descending_refinement_rates",
"schema": "community-qwen235b-boundary-repeat-selection-v4",
"selected_configs": [
"tp4_mns128_mbt16384",
"tp4_mns128_mbt8192",
"tp4_mns64_mbt16384",
"tp4_mns64_mbt8192",
"tp8_mns128_mbt8192",
"tp8_mns64_mbt8192"
],
"selected_slos": [
"linear_8k_primary"
],
"selection": "all_adjacent_pre_registered_slo_feasibility_transitions_plus_between_rate_refinement",
"selections": [
{
"added_refinement_rates": [],
"boundary_rates": [
0.3,
0.35
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.3,
0.35
]
},
"config": "tp4_mns128_mbt16384",
"excluded_unrankable_slos": [],
"repeat_order": [
0.3,
0.35
],
"selected_rates": [
0.3,
0.35
],
"source_relative_order": [
0.35,
0.3
]
},
{
"added_refinement_rates": [],
"boundary_rates": [
0.2,
0.25
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.2,
0.25
]
},
"config": "tp4_mns64_mbt8192",
"excluded_unrankable_slos": [],
"repeat_order": [
0.2,
0.25
],
"selected_rates": [
0.2,
0.25
],
"source_relative_order": [
0.25,
0.2
]
},
{
"added_refinement_rates": [],
"boundary_rates": [
0.2,
0.25
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.2,
0.25
]
},
"config": "tp4_mns128_mbt8192",
"excluded_unrankable_slos": [],
"repeat_order": [
0.2,
0.25
],
"selected_rates": [
0.2,
0.25
],
"source_relative_order": [
0.25,
0.2
]
},
{
"added_refinement_rates": [],
"boundary_rates": [
0.45,
0.5
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.45,
0.5
]
},
"config": "tp8_mns128_mbt8192",
"excluded_unrankable_slos": [],
"repeat_order": [
0.45,
0.5
],
"selected_rates": [
0.45,
0.5
],
"source_relative_order": [
0.5,
0.45
]
},
{
"added_refinement_rates": [],
"boundary_rates": [
0.45,
0.5
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.45,
0.5
]
},
"config": "tp8_mns64_mbt8192",
"excluded_unrankable_slos": [],
"repeat_order": [
0.45,
0.5
],
"selected_rates": [
0.45,
0.5
],
"source_relative_order": [
0.5,
0.45
]
},
{
"added_refinement_rates": [],
"boundary_rates": [
0.3,
0.35
],
"boundary_rates_by_slo": {
"linear_8k_primary": [
0.3,
0.35
]
},
"config": "tp4_mns64_mbt16384",
"excluded_unrankable_slos": [],
"repeat_order": [
0.3,
0.35
],
"selected_rates": [
0.3,
0.35
],
"source_relative_order": [
0.35,
0.3
]
}
],
"unrankable_slo_policy": "exclude_per_config_surfaces_with_no_feasible_load",
"warmup": {
"latencies_discarded": true,
"method": "one_disjoint_request_per_input_length_bin",
"request_count": 6,
"seed": 2026071504
}
},
"frontier_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/frontier-v4-refined-rates-batched-par4/frontier_ranking_frozen.json",
"sha256": "318b3c7742c8371191e2dbd429f0e0e63c1a42f4806a599e018b0548cb9d6336"
},
"model": {
"config_sha256": "702c46d431bb984db9035a1225186bbfdb52c0d19c82104df4a37cd005e0369e",
"path": "/home/admin/cpfs/wjh/models/Qwen/Qwen3-235B-A22B-FP8"
},
"protocol_manifest": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/refinement-v1/protocol/protocol_manifest.json",
"sha256": "695c0500b9a5d51cdd4e39b722e3c9db15f8bed48795454f4d0d6442d24929cb"
},
"repeat_of": {
"first_pass_freeze": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/community_ranking_frozen.json",
"sha256": "5a9c59fc4082db47bec7b59432e9ebe9f907f70baa95a9f0d3480e85f8992ff2"
},
"source_manifest": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v2-boundary-reverse/run_manifest.json",
"sha256": "cf1a722043f42ff8f435e3521d0b2b2892a8ea0922e1e4870ec10fba26c0f998"
}
},
"repository": {
"commit": "344af3a4285ccfdc6e10e540767808907ad81ddf",
"status_porcelain": []
},
"runtime": {
"python": "/usr/bin/python3.12",
"vllm": "/tmp/wjh-frontier-vllm0102-smoke/.venv/bin/vllm"
},
"schema": "community-qwen235b-prefill-fixed-cohort-v1",
"studies": {
"4": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp4/study.json",
"sha256": "ea0d7d31e4a5102cf530e900754debb1d8b18f860748ead88ebf7de57d92faa3"
},
"8": {
"path": "/home/admin/cpfs/wjh/frontier-community-qwen235-smoke-20260715/rankable-fixed-cohort-v1/community-v1-blind/studies/tp8/study.json",
"sha256": "48c4d917b5531f742a05bab3bcd5c880f2eaddf9992b7629e9f48d8e00fd1906"
}
}
}