svd-safety-mis7_swift_remove40 / compression.json
Jeesup's picture
None arm at None budget
d84016f verified
Raw History Blame Contribute Delete
15.2 kB
{
"source": "/workspace/svd_safety/assets/models/Mistral-7B-Instruct-v0.2",
"upstream": "hiahei/Swift-SVD (compression) + AIoT-MLSys-Lab/SVD-LLM utils/LoRA.py (recovery)",
"stage": "finetuned",
"ratio_removed": 0.4,
"internal_ratio_kept": 0.6,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/mis7_swift_remove40/finetune_v/merge.pt",
"mean_rel_err": 0.524213519213455,
"max_rel_err": 0.6444101929664612,
"param_fraction": 0.6003230168269231,
"ranks": {
"0.q_proj": 1383,
"0.k_proj": 581,
"0.v_proj": 529,
"0.o_proj": 1263,
"0.gate_proj": 2016,
"0.up_proj": 2016,
"0.down_proj": 2043,
"1.q_proj": 1388,
"1.k_proj": 591,
"1.v_proj": 557,
"1.o_proj": 1300,
"1.gate_proj": 2016,
"1.up_proj": 2016,
"1.down_proj": 2371,
"2.q_proj": 1131,
"2.k_proj": 448,
"2.v_proj": 439,
"2.o_proj": 1088,
"2.gate_proj": 2016,
"2.up_proj": 2016,
"2.down_proj": 1622,
"3.q_proj": 1128,
"3.k_proj": 446,
"3.v_proj": 435,
"3.o_proj": 1077,
"3.gate_proj": 2016,
"3.up_proj": 2016,
"3.down_proj": 1624,
"4.q_proj": 1126,
"4.k_proj": 445,
"4.v_proj": 439,
"4.o_proj": 1084,
"4.gate_proj": 2016,
"4.up_proj": 2016,
"4.down_proj": 1645,
"5.q_proj": 1126,
"5.k_proj": 445,
"5.v_proj": 439,
"5.o_proj": 1092,
"5.gate_proj": 2016,
"5.up_proj": 2016,
"5.down_proj": 1662,
"6.q_proj": 1127,
"6.k_proj": 446,
"6.v_proj": 440,
"6.o_proj": 1101,
"6.gate_proj": 2016,
"6.up_proj": 2016,
"6.down_proj": 1675,
"7.q_proj": 1126,
"7.k_proj": 446,
"7.v_proj": 441,
"7.o_proj": 1110,
"7.gate_proj": 2016,
"7.up_proj": 2016,
"7.down_proj": 1688,
"8.q_proj": 1122,
"8.k_proj": 444,
"8.v_proj": 441,
"8.o_proj": 1109,
"8.gate_proj": 2016,
"8.up_proj": 2016,
"8.down_proj": 1685,
"9.q_proj": 1121,
"9.k_proj": 444,
"9.v_proj": 442,
"9.o_proj": 1109,
"9.gate_proj": 2016,
"9.up_proj": 2016,
"9.down_proj": 1692,
"10.q_proj": 1122,
"10.k_proj": 444,
"10.v_proj": 442,
"10.o_proj": 1118,
"10.gate_proj": 2016,
"10.up_proj": 2016,
"10.down_proj": 1692,
"11.q_proj": 1120,
"11.k_proj": 444,
"11.v_proj": 443,
"11.o_proj": 1121,
"11.gate_proj": 2016,
"11.up_proj": 2016,
"11.down_proj": 1695,
"12.q_proj": 1127,
"12.k_proj": 447,
"12.v_proj": 447,
"12.o_proj": 1129,
"12.gate_proj": 2016,
"12.up_proj": 2016,
"12.down_proj": 1714,
"13.q_proj": 1123,
"13.k_proj": 446,
"13.v_proj": 445,
"13.o_proj": 1127,
"13.gate_proj": 2016,
"13.up_proj": 2016,
"13.down_proj": 1714,
"14.q_proj": 1132,
"14.k_proj": 451,
"14.v_proj": 453,
"14.o_proj": 1139,
"14.gate_proj": 2016,
"14.up_proj": 2016,
"14.down_proj": 1742,
"15.q_proj": 1145,
"15.k_proj": 458,
"15.v_proj": 461,
"15.o_proj": 1157,
"15.gate_proj": 2016,
"15.up_proj": 2016,
"15.down_proj": 1777,
"16.q_proj": 1151,
"16.k_proj": 461,
"16.v_proj": 463,
"16.o_proj": 1162,
"16.gate_proj": 2016,
"16.up_proj": 2016,
"16.down_proj": 1800,
"17.q_proj": 1174,
"17.k_proj": 473,
"17.v_proj": 476,
"17.o_proj": 1188,
"17.gate_proj": 2016,
"17.up_proj": 2016,
"17.down_proj": 1852,
"18.q_proj": 1174,
"18.k_proj": 472,
"18.v_proj": 476,
"18.o_proj": 1191,
"18.gate_proj": 2016,
"18.up_proj": 2016,
"18.down_proj": 1866,
"19.q_proj": 1158,
"19.k_proj": 465,
"19.v_proj": 469,
"19.o_proj": 1179,
"19.gate_proj": 2016,
"19.up_proj": 2016,
"19.down_proj": 1844,
"20.q_proj": 1130,
"20.k_proj": 449,
"20.v_proj": 454,
"20.o_proj": 1146,
"20.gate_proj": 2016,
"20.up_proj": 2016,
"20.down_proj": 1786,
"21.q_proj": 1123,
"21.k_proj": 444,
"21.v_proj": 451,
"21.o_proj": 1137,
"21.gate_proj": 2016,
"21.up_proj": 2016,
"21.down_proj": 1772,
"22.q_proj": 1112,
"22.k_proj": 437,
"22.v_proj": 445,
"22.o_proj": 1121,
"22.gate_proj": 2016,
"22.up_proj": 2016,
"22.down_proj": 1749,
"23.q_proj": 1112,
"23.k_proj": 438,
"23.v_proj": 444,
"23.o_proj": 1129,
"23.gate_proj": 2016,
"23.up_proj": 2016,
"23.down_proj": 1752,
"24.q_proj": 1112,
"24.k_proj": 437,
"24.v_proj": 445,
"24.o_proj": 1133,
"24.gate_proj": 2016,
"24.up_proj": 2016,
"24.down_proj": 1754,
"25.q_proj": 1114,
"25.k_proj": 439,
"25.v_proj": 449,
"25.o_proj": 1141,
"25.gate_proj": 2016,
"25.up_proj": 2016,
"25.down_proj": 1766,
"26.q_proj": 1119,
"26.k_proj": 441,
"26.v_proj": 452,
"26.o_proj": 1149,
"26.gate_proj": 2016,
"26.up_proj": 2016,
"26.down_proj": 1780,
"27.q_proj": 1123,
"27.k_proj": 444,
"27.v_proj": 454,
"27.o_proj": 1149,
"27.gate_proj": 2016,
"27.up_proj": 2016,
"27.down_proj": 1792,
"28.q_proj": 1139,
"28.k_proj": 452,
"28.v_proj": 463,
"28.o_proj": 1171,
"28.gate_proj": 2016,
"28.up_proj": 2016,
"28.down_proj": 1838,
"29.q_proj": 1148,
"29.k_proj": 457,
"29.v_proj": 469,
"29.o_proj": 1193,
"29.gate_proj": 2016,
"29.up_proj": 2016,
"29.down_proj": 1861,
"30.q_proj": 1165,
"30.k_proj": 466,
"30.v_proj": 481,
"30.o_proj": 1215,
"30.gate_proj": 2016,
"30.up_proj": 2016,
"30.down_proj": 1912,
"31.q_proj": 1263,
"31.k_proj": 519,
"31.v_proj": 536,
"31.o_proj": 1336,
"31.gate_proj": 2016,
"31.up_proj": 2016,
"31.down_proj": 2179
},
"rel_err": {
"0.q_proj": 0.09311120212078094,
"0.k_proj": 0.08253687620162964,
"0.v_proj": 0.3433501720428467,
"0.o_proj": 0.3911000192165375,
"0.gate_proj": 0.3707573413848877,
"0.up_proj": 0.41067802906036377,
"0.down_proj": 0.5313096046447754,
"1.q_proj": 0.16670213639736176,
"1.k_proj": 0.10798212885856628,
"1.v_proj": 0.5118529796600342,
"1.o_proj": 0.5213813781738281,
"1.gate_proj": 0.4368397891521454,
"1.up_proj": 0.4705049991607666,
"1.down_proj": 0.4971780776977539,
"2.q_proj": 0.46672654151916504,
"2.k_proj": 0.46059876680374146,
"2.v_proj": 0.6164757013320923,
"2.o_proj": 0.5800544619560242,
"2.gate_proj": 0.512566328048706,
"2.up_proj": 0.5365164875984192,
"2.down_proj": 0.6227588057518005,
"3.q_proj": 0.4239010214805603,
"3.k_proj": 0.3750090003013611,
"3.v_proj": 0.6444101929664612,
"3.o_proj": 0.6232957243919373,
"3.gate_proj": 0.5219470858573914,
"3.up_proj": 0.551502525806427,
"3.down_proj": 0.6144695281982422,
"4.q_proj": 0.4298405945301056,
"4.k_proj": 0.3984021842479706,
"4.v_proj": 0.6045357584953308,
"4.o_proj": 0.5685608386993408,
"4.gate_proj": 0.523536205291748,
"4.up_proj": 0.5611042976379395,
"4.down_proj": 0.6211814284324646,
"5.q_proj": 0.46155083179473877,
"5.k_proj": 0.4410848319530487,
"5.v_proj": 0.6360407471656799,
"5.o_proj": 0.6054348349571228,
"5.gate_proj": 0.5178220868110657,
"5.up_proj": 0.5632009506225586,
"5.down_proj": 0.6081773638725281,
"6.q_proj": 0.4475008249282837,
"6.k_proj": 0.4137071967124939,
"6.v_proj": 0.6230500340461731,
"6.o_proj": 0.5891891121864319,
"6.gate_proj": 0.5309041738510132,
"6.up_proj": 0.5662814378738403,
"6.down_proj": 0.5977262258529663,
"7.q_proj": 0.427289217710495,
"7.k_proj": 0.38710105419158936,
"7.v_proj": 0.6161573529243469,
"7.o_proj": 0.5648438930511475,
"7.gate_proj": 0.5202533602714539,
"7.up_proj": 0.5593512654304504,
"7.down_proj": 0.592329740524292,
"8.q_proj": 0.43415379524230957,
"8.k_proj": 0.3839348256587982,
"8.v_proj": 0.576714813709259,
"8.o_proj": 0.5553375482559204,
"8.gate_proj": 0.523369312286377,
"8.up_proj": 0.5563741326332092,
"8.down_proj": 0.5858190059661865,
"9.q_proj": 0.42894482612609863,
"9.k_proj": 0.3965420722961426,
"9.v_proj": 0.5610228180885315,
"9.o_proj": 0.5573288202285767,
"9.gate_proj": 0.5195508599281311,
"9.up_proj": 0.5543780326843262,
"9.down_proj": 0.5704837441444397,
"10.q_proj": 0.42824241518974304,
"10.k_proj": 0.401052862405777,
"10.v_proj": 0.5595612525939941,
"10.o_proj": 0.5597419738769531,
"10.gate_proj": 0.5190359354019165,
"10.up_proj": 0.5517367720603943,
"10.down_proj": 0.5768575668334961,
"11.q_proj": 0.4241175949573517,
"11.k_proj": 0.3747548758983612,
"11.v_proj": 0.5366345643997192,
"11.o_proj": 0.5581367611885071,
"11.gate_proj": 0.5184133052825928,
"11.up_proj": 0.5504465103149414,
"11.down_proj": 0.5751660466194153,
"12.q_proj": 0.41172537207603455,
"12.k_proj": 0.3747641444206238,
"12.v_proj": 0.5728425979614258,
"12.o_proj": 0.5404180288314819,
"12.gate_proj": 0.5148531198501587,
"12.up_proj": 0.5455449223518372,
"12.down_proj": 0.5717161893844604,
"13.q_proj": 0.4387548565864563,
"13.k_proj": 0.40697500109672546,
"13.v_proj": 0.5708914399147034,
"13.o_proj": 0.5474628210067749,
"13.gate_proj": 0.5218628644943237,
"13.up_proj": 0.5454148054122925,
"13.down_proj": 0.5676140189170837,
"14.q_proj": 0.4094313979148865,
"14.k_proj": 0.38163262605667114,
"14.v_proj": 0.5798335075378418,
"14.o_proj": 0.5494734048843384,
"14.gate_proj": 0.5187147259712219,
"14.up_proj": 0.5501715540885925,
"14.down_proj": 0.593805193901062,
"15.q_proj": 0.40906858444213867,
"15.k_proj": 0.3624928295612335,
"15.v_proj": 0.6080702543258667,
"15.o_proj": 0.5543991327285767,
"15.gate_proj": 0.5166524648666382,
"15.up_proj": 0.5478196144104004,
"15.down_proj": 0.5904906988143921,
"16.q_proj": 0.4163205027580261,
"16.k_proj": 0.3787270486354828,
"16.v_proj": 0.6027750372886658,
"16.o_proj": 0.5409226417541504,
"16.gate_proj": 0.520693838596344,
"16.up_proj": 0.556447446346283,
"16.down_proj": 0.5904657244682312,
"17.q_proj": 0.4439752697944641,
"17.k_proj": 0.41725632548332214,
"17.v_proj": 0.5800758004188538,
"17.o_proj": 0.5581274628639221,
"17.gate_proj": 0.5324439406394958,
"17.up_proj": 0.5666811466217041,
"17.down_proj": 0.5825478434562683,
"18.q_proj": 0.4187489449977875,
"18.k_proj": 0.3872785270214081,
"18.v_proj": 0.6302375793457031,
"18.o_proj": 0.5343031883239746,
"18.gate_proj": 0.5309059619903564,
"18.up_proj": 0.564786434173584,
"18.down_proj": 0.5998350381851196,
"19.q_proj": 0.4693866968154907,
"19.k_proj": 0.40707793831825256,
"19.v_proj": 0.6421195268630981,
"19.o_proj": 0.5577782392501831,
"19.gate_proj": 0.536626398563385,
"19.up_proj": 0.5659677386283875,
"19.down_proj": 0.6147719621658325,
"20.q_proj": 0.4903647303581238,
"20.k_proj": 0.44427719712257385,
"20.v_proj": 0.6202300190925598,
"20.o_proj": 0.5741354823112488,
"20.gate_proj": 0.5429403185844421,
"20.up_proj": 0.5663204789161682,
"20.down_proj": 0.6225840449333191,
"21.q_proj": 0.5229761004447937,
"21.k_proj": 0.47562214732170105,
"21.v_proj": 0.5926244854927063,
"21.o_proj": 0.611431360244751,
"21.gate_proj": 0.5432093143463135,
"21.up_proj": 0.5681016445159912,
"21.down_proj": 0.6183050274848938,
"22.q_proj": 0.5626530051231384,
"22.k_proj": 0.5008488297462463,
"22.v_proj": 0.5983458757400513,
"22.o_proj": 0.6302191019058228,
"22.gate_proj": 0.5491480231285095,
"22.up_proj": 0.5697852969169617,
"22.down_proj": 0.6164385676383972,
"23.q_proj": 0.551480233669281,
"23.k_proj": 0.5012710690498352,
"23.v_proj": 0.6238268613815308,
"23.o_proj": 0.6080763339996338,
"23.gate_proj": 0.5521300435066223,
"23.up_proj": 0.5716087222099304,
"23.down_proj": 0.6099771857261658,
"24.q_proj": 0.5572198629379272,
"24.k_proj": 0.5183809399604797,
"24.v_proj": 0.6330755352973938,
"24.o_proj": 0.6300069689750671,
"24.gate_proj": 0.5530964136123657,
"24.up_proj": 0.5734592080116272,
"24.down_proj": 0.600332498550415,
"25.q_proj": 0.5632444620132446,
"25.k_proj": 0.5026987791061401,
"25.v_proj": 0.6118474006652832,
"25.o_proj": 0.6110818982124329,
"25.gate_proj": 0.5554454922676086,
"25.up_proj": 0.57483971118927,
"25.down_proj": 0.588827908039093,
"26.q_proj": 0.558449923992157,
"26.k_proj": 0.4919636845588684,
"26.v_proj": 0.5971678495407104,
"26.o_proj": 0.5927439332008362,
"26.gate_proj": 0.5562448501586914,
"26.up_proj": 0.5749430060386658,
"26.down_proj": 0.5749134421348572,
"27.q_proj": 0.5691151022911072,
"27.k_proj": 0.5340715050697327,
"27.v_proj": 0.6143185496330261,
"27.o_proj": 0.5845725536346436,
"27.gate_proj": 0.5597459673881531,
"27.up_proj": 0.5763874650001526,
"27.down_proj": 0.5742335319519043,
"28.q_proj": 0.5281823873519897,
"28.k_proj": 0.4590943157672882,
"28.v_proj": 0.5685400366783142,
"28.o_proj": 0.5630350112915039,
"28.gate_proj": 0.5594995617866516,
"28.up_proj": 0.5738884806632996,
"28.down_proj": 0.5813770294189453,
"29.q_proj": 0.5108082890510559,
"29.k_proj": 0.46057870984077454,
"29.v_proj": 0.6014853715896606,
"29.o_proj": 0.5444827079772949,
"29.gate_proj": 0.5542632937431335,
"29.up_proj": 0.5680719017982483,
"29.down_proj": 0.5874636769294739,
"30.q_proj": 0.4883614480495453,
"30.k_proj": 0.4195691645145416,
"30.v_proj": 0.535009503364563,
"30.o_proj": 0.5329458713531494,
"30.gate_proj": 0.5540768504142761,
"30.up_proj": 0.5595806241035461,
"30.down_proj": 0.5732670426368713,
"31.q_proj": 0.4402003884315491,
"31.k_proj": 0.3733157217502594,
"31.v_proj": 0.48748257756233215,
"31.o_proj": 0.45593205094337463,
"31.gate_proj": 0.5609548091888428,
"31.up_proj": 0.5358024835586548,
"31.down_proj": 0.5254908800125122
},
"step": "swift_svd_dynamic_finetuned",
"compression": {
"upstream": "hiahei/Swift-SVD",
"step": "swift_svd_dynamic",
"alloc": "dynamic",
"alpha": 0.6,
"nsamples": 256,
"seqlen": 2048,
"dataset": "WikiText2",
"seed": 42,
"gqa_budget": "shape",
"param_fraction": 0.6003230168269231,
"mean_rel_err": 0.523012746524598,
"max_rel_err": 0.6436969041824341
},
"factorisation": {
"max_rel_err": 0.001442187582142651,
"factor_params": 4189847552,
"dense_params": 6979321856
},
"finetune": {
"protocol": "svdllm_stage2_u_then_v",
"script": "gitclones/SVD-LLM/utils/LoRA.py",
"data": "alpaca-cleaned",
"epochs_per_half": 2,
"learning_rate": 0.0001,
"batch_size": 64,
"micro_batch_size": 4,
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"cutoff_len": 256,
"val_set_size": 2000,
"max_train_samples": "all",
"precision": "bf16 (upstream hard-codes fp16; see run_with_torch_compat)"
}
}