svd-safety-l3_swift_remove50 / compression.json
Jeesup's picture
None arm at None budget
e0fb00a verified
Raw History Blame Contribute Delete
15.1 kB
{
"source": "/workspace/svd_safety/assets/models/Llama-3-8B-Instruct",
"upstream": "hiahei/Swift-SVD (compression) + AIoT-MLSys-Lab/SVD-LLM utils/LoRA.py (recovery)",
"stage": "finetuned",
"ratio_removed": 0.5,
"internal_ratio_kept": 0.5,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/l3_swift_remove50/finetune_v/merge.pt",
"mean_rel_err": 0.5984087092136698,
"max_rel_err": 0.7104170918464661,
"param_fraction": 0.5002723106971154,
"ranks": {
"0.q_proj": 1153,
"0.k_proj": 486,
"0.v_proj": 455,
"0.o_proj": 1097,
"0.gate_proj": 1680,
"0.up_proj": 1680,
"0.down_proj": 1764,
"1.q_proj": 1001,
"1.k_proj": 405,
"1.v_proj": 389,
"1.o_proj": 959,
"1.gate_proj": 1680,
"1.up_proj": 1680,
"1.down_proj": 1618,
"2.q_proj": 993,
"2.k_proj": 401,
"2.v_proj": 396,
"2.o_proj": 962,
"2.gate_proj": 1680,
"2.up_proj": 1680,
"2.down_proj": 1484,
"3.q_proj": 979,
"3.k_proj": 394,
"3.v_proj": 392,
"3.o_proj": 954,
"3.gate_proj": 1680,
"3.up_proj": 1680,
"3.down_proj": 1474,
"4.q_proj": 987,
"4.k_proj": 399,
"4.v_proj": 396,
"4.o_proj": 970,
"4.gate_proj": 1680,
"4.up_proj": 1680,
"4.down_proj": 1506,
"5.q_proj": 979,
"5.k_proj": 395,
"5.v_proj": 391,
"5.o_proj": 966,
"5.gate_proj": 1680,
"5.up_proj": 1680,
"5.down_proj": 1501,
"6.q_proj": 970,
"6.k_proj": 390,
"6.v_proj": 388,
"6.o_proj": 963,
"6.gate_proj": 1680,
"6.up_proj": 1680,
"6.down_proj": 1490,
"7.q_proj": 969,
"7.k_proj": 390,
"7.v_proj": 388,
"7.o_proj": 967,
"7.gate_proj": 1680,
"7.up_proj": 1680,
"7.down_proj": 1489,
"8.q_proj": 959,
"8.k_proj": 385,
"8.v_proj": 383,
"8.o_proj": 958,
"8.gate_proj": 1680,
"8.up_proj": 1680,
"8.down_proj": 1472,
"9.q_proj": 957,
"9.k_proj": 384,
"9.v_proj": 384,
"9.o_proj": 957,
"9.gate_proj": 1680,
"9.up_proj": 1680,
"9.down_proj": 1471,
"10.q_proj": 949,
"10.k_proj": 379,
"10.v_proj": 378,
"10.o_proj": 949,
"10.gate_proj": 1680,
"10.up_proj": 1680,
"10.down_proj": 1457,
"11.q_proj": 955,
"11.k_proj": 382,
"11.v_proj": 383,
"11.o_proj": 956,
"11.gate_proj": 1680,
"11.up_proj": 1680,
"11.down_proj": 1471,
"12.q_proj": 960,
"12.k_proj": 384,
"12.v_proj": 385,
"12.o_proj": 963,
"12.gate_proj": 1680,
"12.up_proj": 1680,
"12.down_proj": 1480,
"13.q_proj": 956,
"13.k_proj": 384,
"13.v_proj": 384,
"13.o_proj": 962,
"13.gate_proj": 1680,
"13.up_proj": 1680,
"13.down_proj": 1479,
"14.q_proj": 956,
"14.k_proj": 384,
"14.v_proj": 384,
"14.o_proj": 962,
"14.gate_proj": 1680,
"14.up_proj": 1680,
"14.down_proj": 1481,
"15.q_proj": 959,
"15.k_proj": 384,
"15.v_proj": 384,
"15.o_proj": 963,
"15.gate_proj": 1680,
"15.up_proj": 1680,
"15.down_proj": 1490,
"16.q_proj": 957,
"16.k_proj": 383,
"16.v_proj": 382,
"16.o_proj": 963,
"16.gate_proj": 1680,
"16.up_proj": 1680,
"16.down_proj": 1488,
"17.q_proj": 946,
"17.k_proj": 376,
"17.v_proj": 376,
"17.o_proj": 949,
"17.gate_proj": 1680,
"17.up_proj": 1680,
"17.down_proj": 1468,
"18.q_proj": 936,
"18.k_proj": 371,
"18.v_proj": 372,
"18.o_proj": 935,
"18.gate_proj": 1680,
"18.up_proj": 1680,
"18.down_proj": 1450,
"19.q_proj": 931,
"19.k_proj": 368,
"19.v_proj": 370,
"19.o_proj": 927,
"19.gate_proj": 1680,
"19.up_proj": 1680,
"19.down_proj": 1440,
"20.q_proj": 928,
"20.k_proj": 367,
"20.v_proj": 369,
"20.o_proj": 925,
"20.gate_proj": 1680,
"20.up_proj": 1680,
"20.down_proj": 1436,
"21.q_proj": 931,
"21.k_proj": 368,
"21.v_proj": 370,
"21.o_proj": 935,
"21.gate_proj": 1680,
"21.up_proj": 1680,
"21.down_proj": 1444,
"22.q_proj": 925,
"22.k_proj": 365,
"22.v_proj": 368,
"22.o_proj": 925,
"22.gate_proj": 1680,
"22.up_proj": 1680,
"22.down_proj": 1436,
"23.q_proj": 924,
"23.k_proj": 364,
"23.v_proj": 367,
"23.o_proj": 923,
"23.gate_proj": 1680,
"23.up_proj": 1680,
"23.down_proj": 1433,
"24.q_proj": 923,
"24.k_proj": 363,
"24.v_proj": 368,
"24.o_proj": 926,
"24.gate_proj": 1680,
"24.up_proj": 1680,
"24.down_proj": 1431,
"25.q_proj": 925,
"25.k_proj": 364,
"25.v_proj": 369,
"25.o_proj": 930,
"25.gate_proj": 1680,
"25.up_proj": 1680,
"25.down_proj": 1434,
"26.q_proj": 924,
"26.k_proj": 364,
"26.v_proj": 368,
"26.o_proj": 939,
"26.gate_proj": 1680,
"26.up_proj": 1680,
"26.down_proj": 1440,
"27.q_proj": 925,
"27.k_proj": 364,
"27.v_proj": 371,
"27.o_proj": 944,
"27.gate_proj": 1680,
"27.up_proj": 1680,
"27.down_proj": 1448,
"28.q_proj": 928,
"28.k_proj": 366,
"28.v_proj": 372,
"28.o_proj": 945,
"28.gate_proj": 1680,
"28.up_proj": 1680,
"28.down_proj": 1462,
"29.q_proj": 933,
"29.k_proj": 369,
"29.v_proj": 376,
"29.o_proj": 961,
"29.gate_proj": 1680,
"29.up_proj": 1680,
"29.down_proj": 1482,
"30.q_proj": 944,
"30.k_proj": 374,
"30.v_proj": 382,
"30.o_proj": 978,
"30.gate_proj": 1680,
"30.up_proj": 1680,
"30.down_proj": 1527,
"31.q_proj": 1058,
"31.k_proj": 436,
"31.v_proj": 448,
"31.o_proj": 1107,
"31.gate_proj": 1680,
"31.up_proj": 1680,
"31.down_proj": 1830
},
"rel_err": {
"0.q_proj": 0.1555257886648178,
"0.k_proj": 0.11019963026046753,
"0.v_proj": 0.6514365673065186,
"0.o_proj": 0.558626651763916,
"0.gate_proj": 0.5772461295127869,
"0.up_proj": 0.6381028890609741,
"0.down_proj": 0.6298868060112,
"1.q_proj": 0.3837055265903473,
"1.k_proj": 0.340504914522171,
"1.v_proj": 0.6825932860374451,
"1.o_proj": 0.6070690155029297,
"1.gate_proj": 0.5934946537017822,
"1.up_proj": 0.6246401071548462,
"1.down_proj": 0.6702373623847961,
"2.q_proj": 0.5115190148353577,
"2.k_proj": 0.46816286444664,
"2.v_proj": 0.6202782988548279,
"2.o_proj": 0.5873217582702637,
"2.gate_proj": 0.6200411319732666,
"2.up_proj": 0.6416133642196655,
"2.down_proj": 0.6713889241218567,
"3.q_proj": 0.5343612432479858,
"3.k_proj": 0.48675867915153503,
"3.v_proj": 0.6520168781280518,
"3.o_proj": 0.6289046406745911,
"3.gate_proj": 0.6113771796226501,
"3.up_proj": 0.649507462978363,
"3.down_proj": 0.664247989654541,
"4.q_proj": 0.5104104280471802,
"4.k_proj": 0.4806947410106659,
"4.v_proj": 0.6609205603599548,
"4.o_proj": 0.6258541941642761,
"4.gate_proj": 0.6113108992576599,
"4.up_proj": 0.6584666967391968,
"4.down_proj": 0.6677454113960266,
"5.q_proj": 0.4832381010055542,
"5.k_proj": 0.49522989988327026,
"5.v_proj": 0.6369020938873291,
"5.o_proj": 0.6074650883674622,
"5.gate_proj": 0.6114932298660278,
"5.up_proj": 0.6566051840782166,
"5.down_proj": 0.663334310054779,
"6.q_proj": 0.4991042912006378,
"6.k_proj": 0.4993811845779419,
"6.v_proj": 0.6485827565193176,
"6.o_proj": 0.6202296018600464,
"6.gate_proj": 0.6068925857543945,
"6.up_proj": 0.6527253985404968,
"6.down_proj": 0.6605090498924255,
"7.q_proj": 0.4872052073478699,
"7.k_proj": 0.485134094953537,
"7.v_proj": 0.6055662631988525,
"7.o_proj": 0.6058576107025146,
"7.gate_proj": 0.6032496094703674,
"7.up_proj": 0.6502962112426758,
"7.down_proj": 0.641866147518158,
"8.q_proj": 0.4624485671520233,
"8.k_proj": 0.46426060795783997,
"8.v_proj": 0.6231876611709595,
"8.o_proj": 0.5961927771568298,
"8.gate_proj": 0.5874994993209839,
"8.up_proj": 0.6398731470108032,
"8.down_proj": 0.6440942883491516,
"9.q_proj": 0.4597213566303253,
"9.k_proj": 0.46052107214927673,
"9.v_proj": 0.572304368019104,
"9.o_proj": 0.6035570502281189,
"9.gate_proj": 0.5833413600921631,
"9.up_proj": 0.6350976824760437,
"9.down_proj": 0.6436678767204285,
"10.q_proj": 0.46521714329719543,
"10.k_proj": 0.478322833776474,
"10.v_proj": 0.5912260413169861,
"10.o_proj": 0.6270232796669006,
"10.gate_proj": 0.5851974487304688,
"10.up_proj": 0.6311303377151489,
"10.down_proj": 0.6381176710128784,
"11.q_proj": 0.47614750266075134,
"11.k_proj": 0.4854636788368225,
"11.v_proj": 0.5704461932182312,
"11.o_proj": 0.60197913646698,
"11.gate_proj": 0.5863650441169739,
"11.up_proj": 0.6343603134155273,
"11.down_proj": 0.6290261745452881,
"12.q_proj": 0.4840157628059387,
"12.k_proj": 0.4766708314418793,
"12.v_proj": 0.6087822318077087,
"12.o_proj": 0.6030346751213074,
"12.gate_proj": 0.5890772342681885,
"12.up_proj": 0.6384384036064148,
"12.down_proj": 0.6522207856178284,
"13.q_proj": 0.43680158257484436,
"13.k_proj": 0.4298385679721832,
"13.v_proj": 0.6261775493621826,
"13.o_proj": 0.5952807664871216,
"13.gate_proj": 0.5872403383255005,
"13.up_proj": 0.6357120275497437,
"13.down_proj": 0.6553712487220764,
"14.q_proj": 0.45404571294784546,
"14.k_proj": 0.4589211046695709,
"14.v_proj": 0.6460106372833252,
"14.o_proj": 0.6136132478713989,
"14.gate_proj": 0.5913381576538086,
"14.up_proj": 0.647808849811554,
"14.down_proj": 0.6647049784660339,
"15.q_proj": 0.45149752497673035,
"15.k_proj": 0.4706929326057434,
"15.v_proj": 0.6710425615310669,
"15.o_proj": 0.6365965604782104,
"15.gate_proj": 0.5898893475532532,
"15.up_proj": 0.6483412384986877,
"15.down_proj": 0.6726868748664856,
"16.q_proj": 0.49306097626686096,
"16.k_proj": 0.4802592396736145,
"16.v_proj": 0.6781746745109558,
"16.o_proj": 0.6331717371940613,
"16.gate_proj": 0.5992128849029541,
"16.up_proj": 0.6493324637413025,
"16.down_proj": 0.6702439785003662,
"17.q_proj": 0.4995686411857605,
"17.k_proj": 0.49193471670150757,
"17.v_proj": 0.6817148923873901,
"17.o_proj": 0.6546281576156616,
"17.gate_proj": 0.6066341996192932,
"17.up_proj": 0.6509790420532227,
"17.down_proj": 0.6842725276947021,
"18.q_proj": 0.5473029017448425,
"18.k_proj": 0.5079021453857422,
"18.v_proj": 0.6422337293624878,
"18.o_proj": 0.6411892771720886,
"18.gate_proj": 0.6166319251060486,
"18.up_proj": 0.6503247022628784,
"18.down_proj": 0.6794316172599792,
"19.q_proj": 0.5715011358261108,
"19.k_proj": 0.5608424544334412,
"19.v_proj": 0.6600127816200256,
"19.o_proj": 0.6745914220809937,
"19.gate_proj": 0.6198409795761108,
"19.up_proj": 0.647895872592926,
"19.down_proj": 0.6834138035774231,
"20.q_proj": 0.568836510181427,
"20.k_proj": 0.5440749526023865,
"20.v_proj": 0.6928178071975708,
"20.o_proj": 0.6790831685066223,
"20.gate_proj": 0.6240135431289673,
"20.up_proj": 0.648634672164917,
"20.down_proj": 0.6841005086898804,
"21.q_proj": 0.5723114609718323,
"21.k_proj": 0.5470589995384216,
"21.v_proj": 0.6950841546058655,
"21.o_proj": 0.6883072257041931,
"21.gate_proj": 0.6248385906219482,
"21.up_proj": 0.6486678123474121,
"21.down_proj": 0.6916183233261108,
"22.q_proj": 0.5811800956726074,
"22.k_proj": 0.5494594573974609,
"22.v_proj": 0.6604698896408081,
"22.o_proj": 0.6805965304374695,
"22.gate_proj": 0.6257846355438232,
"22.up_proj": 0.6461135745048523,
"22.down_proj": 0.6894308924674988,
"23.q_proj": 0.6023896336555481,
"23.k_proj": 0.5706185102462769,
"23.v_proj": 0.6682502031326294,
"23.o_proj": 0.6862603425979614,
"23.gate_proj": 0.6281367540359497,
"23.up_proj": 0.6450577974319458,
"23.down_proj": 0.6818817257881165,
"24.q_proj": 0.5854701995849609,
"24.k_proj": 0.5466554164886475,
"24.v_proj": 0.6416366100311279,
"24.o_proj": 0.6976205110549927,
"24.gate_proj": 0.6292465925216675,
"24.up_proj": 0.6437935829162598,
"24.down_proj": 0.6720544099807739,
"25.q_proj": 0.5890008211135864,
"25.k_proj": 0.573958694934845,
"25.v_proj": 0.6611844897270203,
"25.o_proj": 0.6952077150344849,
"25.gate_proj": 0.6287244558334351,
"25.up_proj": 0.6435160636901855,
"25.down_proj": 0.6599352359771729,
"26.q_proj": 0.5816289186477661,
"26.k_proj": 0.5328107476234436,
"26.v_proj": 0.7104170918464661,
"26.o_proj": 0.6926466822624207,
"26.gate_proj": 0.6278991103172302,
"26.up_proj": 0.6444916129112244,
"26.down_proj": 0.6558243036270142,
"27.q_proj": 0.5827873945236206,
"27.k_proj": 0.5452601313591003,
"27.v_proj": 0.6339237093925476,
"27.o_proj": 0.6614390015602112,
"27.gate_proj": 0.6244891881942749,
"27.up_proj": 0.6438120603561401,
"27.down_proj": 0.649568498134613,
"28.q_proj": 0.5894160270690918,
"28.k_proj": 0.5554025173187256,
"28.v_proj": 0.7103909254074097,
"28.o_proj": 0.651461124420166,
"28.gate_proj": 0.6244070529937744,
"28.up_proj": 0.6442367434501648,
"28.down_proj": 0.6564620733261108,
"29.q_proj": 0.6010609269142151,
"29.k_proj": 0.5750370621681213,
"29.v_proj": 0.6888416409492493,
"29.o_proj": 0.6581999063491821,
"29.gate_proj": 0.6179629564285278,
"29.up_proj": 0.639843761920929,
"29.down_proj": 0.6715606451034546,
"30.q_proj": 0.48527583479881287,
"30.k_proj": 0.4483523666858673,
"30.v_proj": 0.6323948502540588,
"30.o_proj": 0.6071786284446716,
"30.gate_proj": 0.6107315421104431,
"30.up_proj": 0.6275578141212463,
"30.down_proj": 0.6769006848335266,
"31.q_proj": 0.5020249485969543,
"31.k_proj": 0.45986607670783997,
"31.v_proj": 0.5856180787086487,
"31.o_proj": 0.606295108795166,
"31.gate_proj": 0.6043108105659485,
"31.up_proj": 0.6049404740333557,
"31.down_proj": 0.632754921913147
},
"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": 0.5002723106971154,
"mean_rel_err": 0.5978308144424643,
"max_rel_err": 0.7103155851364136
},
"factorisation": {
"max_rel_err": 0.001514429459348321,
"factor_params": 3491561472,
"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)"
}
}