svd-safety-mis7_swift_remove30 / compression.json
Jeesup's picture
None arm at None budget
8a8a4b5 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": 0.3,
"internal_ratio_kept": 0.7,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/mis7_swift_remove30/finetune_v/merge.pt",
"mean_rel_err": 0.46171893551945686,
"max_rel_err": 0.5801312327384949,
"param_fraction": 0.7003361628605769,
"ranks": {
"0.q_proj": 1613,
"0.k_proj": 679,
"0.v_proj": 617,
"0.o_proj": 1473,
"0.gate_proj": 2352,
"0.up_proj": 2352,
"0.down_proj": 2382,
"1.q_proj": 1619,
"1.k_proj": 690,
"1.v_proj": 650,
"1.o_proj": 1517,
"1.gate_proj": 2352,
"1.up_proj": 2352,
"1.down_proj": 2765,
"2.q_proj": 1319,
"2.k_proj": 523,
"2.v_proj": 513,
"2.o_proj": 1270,
"2.gate_proj": 2352,
"2.up_proj": 2352,
"2.down_proj": 1892,
"3.q_proj": 1316,
"3.k_proj": 520,
"3.v_proj": 508,
"3.o_proj": 1257,
"3.gate_proj": 2352,
"3.up_proj": 2352,
"3.down_proj": 1895,
"4.q_proj": 1314,
"4.k_proj": 520,
"4.v_proj": 512,
"4.o_proj": 1264,
"4.gate_proj": 2352,
"4.up_proj": 2352,
"4.down_proj": 1919,
"5.q_proj": 1313,
"5.k_proj": 520,
"5.v_proj": 513,
"5.o_proj": 1274,
"5.gate_proj": 2352,
"5.up_proj": 2352,
"5.down_proj": 1938,
"6.q_proj": 1315,
"6.k_proj": 522,
"6.v_proj": 514,
"6.o_proj": 1285,
"6.gate_proj": 2352,
"6.up_proj": 2352,
"6.down_proj": 1953,
"7.q_proj": 1313,
"7.k_proj": 522,
"7.v_proj": 515,
"7.o_proj": 1295,
"7.gate_proj": 2352,
"7.up_proj": 2352,
"7.down_proj": 1969,
"8.q_proj": 1310,
"8.k_proj": 519,
"8.v_proj": 515,
"8.o_proj": 1294,
"8.gate_proj": 2352,
"8.up_proj": 2352,
"8.down_proj": 1965,
"9.q_proj": 1308,
"9.k_proj": 518,
"9.v_proj": 516,
"9.o_proj": 1294,
"9.gate_proj": 2352,
"9.up_proj": 2352,
"9.down_proj": 1973,
"10.q_proj": 1309,
"10.k_proj": 518,
"10.v_proj": 516,
"10.o_proj": 1305,
"10.gate_proj": 2352,
"10.up_proj": 2352,
"10.down_proj": 1974,
"11.q_proj": 1307,
"11.k_proj": 518,
"11.v_proj": 518,
"11.o_proj": 1308,
"11.gate_proj": 2352,
"11.up_proj": 2352,
"11.down_proj": 1977,
"12.q_proj": 1315,
"12.k_proj": 523,
"12.v_proj": 522,
"12.o_proj": 1318,
"12.gate_proj": 2352,
"12.up_proj": 2352,
"12.down_proj": 1999,
"13.q_proj": 1311,
"13.k_proj": 521,
"13.v_proj": 520,
"13.o_proj": 1315,
"13.gate_proj": 2352,
"13.up_proj": 2352,
"13.down_proj": 1999,
"14.q_proj": 1321,
"14.k_proj": 526,
"14.v_proj": 529,
"14.o_proj": 1328,
"14.gate_proj": 2352,
"14.up_proj": 2352,
"14.down_proj": 2032,
"15.q_proj": 1336,
"15.k_proj": 535,
"15.v_proj": 538,
"15.o_proj": 1350,
"15.gate_proj": 2352,
"15.up_proj": 2352,
"15.down_proj": 2073,
"16.q_proj": 1343,
"16.k_proj": 538,
"16.v_proj": 540,
"16.o_proj": 1355,
"16.gate_proj": 2352,
"16.up_proj": 2352,
"16.down_proj": 2099,
"17.q_proj": 1370,
"17.k_proj": 552,
"17.v_proj": 555,
"17.o_proj": 1386,
"17.gate_proj": 2352,
"17.up_proj": 2352,
"17.down_proj": 2160,
"18.q_proj": 1369,
"18.k_proj": 551,
"18.v_proj": 556,
"18.o_proj": 1390,
"18.gate_proj": 2352,
"18.up_proj": 2352,
"18.down_proj": 2176,
"19.q_proj": 1351,
"19.k_proj": 542,
"19.v_proj": 547,
"19.o_proj": 1376,
"19.gate_proj": 2352,
"19.up_proj": 2352,
"19.down_proj": 2151,
"20.q_proj": 1318,
"20.k_proj": 524,
"20.v_proj": 530,
"20.o_proj": 1337,
"20.gate_proj": 2352,
"20.up_proj": 2352,
"20.down_proj": 2083,
"21.q_proj": 1310,
"21.k_proj": 518,
"21.v_proj": 526,
"21.o_proj": 1325,
"21.gate_proj": 2352,
"21.up_proj": 2352,
"21.down_proj": 2067,
"22.q_proj": 1297,
"22.k_proj": 510,
"22.v_proj": 519,
"22.o_proj": 1308,
"22.gate_proj": 2352,
"22.up_proj": 2352,
"22.down_proj": 2040,
"23.q_proj": 1298,
"23.k_proj": 511,
"23.v_proj": 519,
"23.o_proj": 1317,
"23.gate_proj": 2352,
"23.up_proj": 2352,
"23.down_proj": 2044,
"24.q_proj": 1297,
"24.k_proj": 510,
"24.v_proj": 519,
"24.o_proj": 1321,
"24.gate_proj": 2352,
"24.up_proj": 2352,
"24.down_proj": 2045,
"25.q_proj": 1300,
"25.k_proj": 512,
"25.v_proj": 523,
"25.o_proj": 1330,
"25.gate_proj": 2352,
"25.up_proj": 2352,
"25.down_proj": 2059,
"26.q_proj": 1306,
"26.k_proj": 515,
"26.v_proj": 528,
"26.o_proj": 1340,
"26.gate_proj": 2352,
"26.up_proj": 2352,
"26.down_proj": 2076,
"27.q_proj": 1311,
"27.k_proj": 518,
"27.v_proj": 530,
"27.o_proj": 1341,
"27.gate_proj": 2352,
"27.up_proj": 2352,
"27.down_proj": 2090,
"28.q_proj": 1328,
"28.k_proj": 528,
"28.v_proj": 541,
"28.o_proj": 1366,
"28.gate_proj": 2352,
"28.up_proj": 2352,
"28.down_proj": 2144,
"29.q_proj": 1339,
"29.k_proj": 533,
"29.v_proj": 548,
"29.o_proj": 1392,
"29.gate_proj": 2352,
"29.up_proj": 2352,
"29.down_proj": 2170,
"30.q_proj": 1359,
"30.k_proj": 543,
"30.v_proj": 561,
"30.o_proj": 1418,
"30.gate_proj": 2352,
"30.up_proj": 2352,
"30.down_proj": 2230,
"31.q_proj": 1473,
"31.k_proj": 605,
"31.v_proj": 626,
"31.o_proj": 1559,
"31.gate_proj": 2352,
"31.up_proj": 2352,
"31.down_proj": 2541
},
"rel_err": {
"0.q_proj": 0.08783026784658432,
"0.k_proj": 0.076280377805233,
"0.v_proj": 0.28279638290405273,
"0.o_proj": 0.3450925052165985,
"0.gate_proj": 0.3139413893222809,
"0.up_proj": 0.34659841656684875,
"0.down_proj": 0.4610687792301178,
"1.q_proj": 0.14944231510162354,
"1.k_proj": 0.08638541400432587,
"1.v_proj": 0.42672011256217957,
"1.o_proj": 0.4614779055118561,
"1.gate_proj": 0.37372663617134094,
"1.up_proj": 0.4030474126338959,
"1.down_proj": 0.4160686731338501,
"2.q_proj": 0.40826842188835144,
"2.k_proj": 0.38183024525642395,
"2.v_proj": 0.5488172769546509,
"2.o_proj": 0.5261715054512024,
"2.gate_proj": 0.44494783878326416,
"2.up_proj": 0.46598997712135315,
"2.down_proj": 0.5656055808067322,
"3.q_proj": 0.37075719237327576,
"3.k_proj": 0.3103460371494293,
"3.v_proj": 0.5801312327384949,
"3.o_proj": 0.570527195930481,
"3.gate_proj": 0.45506104826927185,
"3.up_proj": 0.48117780685424805,
"3.down_proj": 0.5562737584114075,
"4.q_proj": 0.3778054118156433,
"4.k_proj": 0.33330559730529785,
"4.v_proj": 0.5392099618911743,
"4.o_proj": 0.5162532329559326,
"4.gate_proj": 0.4575117528438568,
"4.up_proj": 0.49015605449676514,
"4.down_proj": 0.5629571676254272,
"5.q_proj": 0.4073948264122009,
"5.k_proj": 0.3724959194660187,
"5.v_proj": 0.5715569853782654,
"5.o_proj": 0.549954354763031,
"5.gate_proj": 0.45217984914779663,
"5.up_proj": 0.49232661724090576,
"5.down_proj": 0.5492331981658936,
"6.q_proj": 0.39479202032089233,
"6.k_proj": 0.3444615304470062,
"6.v_proj": 0.5562271475791931,
"6.o_proj": 0.5352969169616699,
"6.gate_proj": 0.46358194947242737,
"6.up_proj": 0.49562251567840576,
"6.down_proj": 0.5394559502601624,
"7.q_proj": 0.3770744204521179,
"7.k_proj": 0.32297107577323914,
"7.v_proj": 0.5488503575325012,
"7.o_proj": 0.5085676908493042,
"7.gate_proj": 0.4547872841358185,
"7.up_proj": 0.48921388387680054,
"7.down_proj": 0.5321547985076904,
"8.q_proj": 0.3834538757801056,
"8.k_proj": 0.32134145498275757,
"8.v_proj": 0.5088310837745667,
"8.o_proj": 0.4981762766838074,
"8.gate_proj": 0.4569913148880005,
"8.up_proj": 0.48666849732398987,
"8.down_proj": 0.5265653729438782,
"9.q_proj": 0.3789723515510559,
"9.k_proj": 0.3307971656322479,
"9.v_proj": 0.4929361045360565,
"9.o_proj": 0.5009838342666626,
"9.gate_proj": 0.453980952501297,
"9.up_proj": 0.4856830835342407,
"9.down_proj": 0.5116001963615417,
"10.q_proj": 0.3785707354545593,
"10.k_proj": 0.3373726010322571,
"10.v_proj": 0.4934564232826233,
"10.o_proj": 0.5008266568183899,
"10.gate_proj": 0.4539565443992615,
"10.up_proj": 0.48363447189331055,
"10.down_proj": 0.5184407234191895,
"11.q_proj": 0.3743845522403717,
"11.k_proj": 0.3172908425331116,
"11.v_proj": 0.465146541595459,
"11.o_proj": 0.5011467337608337,
"11.gate_proj": 0.4532534182071686,
"11.up_proj": 0.48237940669059753,
"11.down_proj": 0.516564130783081,
"12.q_proj": 0.36303064227104187,
"12.k_proj": 0.30660074949264526,
"12.v_proj": 0.5052470564842224,
"12.o_proj": 0.4829132556915283,
"12.gate_proj": 0.44975727796554565,
"12.up_proj": 0.47766903042793274,
"12.down_proj": 0.5123953223228455,
"13.q_proj": 0.38796108961105347,
"13.k_proj": 0.3449362516403198,
"13.v_proj": 0.5052615404129028,
"13.o_proj": 0.48829060792922974,
"13.gate_proj": 0.4554385840892792,
"13.up_proj": 0.47726166248321533,
"13.down_proj": 0.5091195106506348,
"14.q_proj": 0.36075472831726074,
"14.k_proj": 0.32303425669670105,
"14.v_proj": 0.5109939575195312,
"14.o_proj": 0.49242928624153137,
"14.gate_proj": 0.4523519277572632,
"14.up_proj": 0.48172876238822937,
"14.down_proj": 0.532478928565979,
"15.q_proj": 0.35985133051872253,
"15.k_proj": 0.30060359835624695,
"15.v_proj": 0.5368859767913818,
"15.o_proj": 0.49520623683929443,
"15.gate_proj": 0.44976216554641724,
"15.up_proj": 0.47884488105773926,
"15.down_proj": 0.527091383934021,
"16.q_proj": 0.3671673536300659,
"16.k_proj": 0.319665789604187,
"16.v_proj": 0.5345363020896912,
"16.o_proj": 0.48225852847099304,
"16.gate_proj": 0.4529358446598053,
"16.up_proj": 0.486669659614563,
"16.down_proj": 0.5261300206184387,
"17.q_proj": 0.39087575674057007,
"17.k_proj": 0.3505445420742035,
"17.v_proj": 0.5095366835594177,
"17.o_proj": 0.4995217025279999,
"17.gate_proj": 0.46473342180252075,
"17.up_proj": 0.496961385011673,
"17.down_proj": 0.5169269442558289,
"18.q_proj": 0.3686104118824005,
"18.k_proj": 0.3233947157859802,
"18.v_proj": 0.5597348213195801,
"18.o_proj": 0.47679799795150757,
"18.gate_proj": 0.4632626473903656,
"18.up_proj": 0.49456414580345154,
"18.down_proj": 0.5329453349113464,
"19.q_proj": 0.4161927103996277,
"19.k_proj": 0.3401549458503723,
"19.v_proj": 0.5778656601905823,
"19.o_proj": 0.49888941645622253,
"19.gate_proj": 0.4682750403881073,
"19.up_proj": 0.4958019256591797,
"19.down_proj": 0.548184871673584,
"20.q_proj": 0.438224732875824,
"20.k_proj": 0.380808025598526,
"20.v_proj": 0.5477811098098755,
"20.o_proj": 0.5176671147346497,
"20.gate_proj": 0.4731369912624359,
"20.up_proj": 0.49586376547813416,
"20.down_proj": 0.5577734112739563,
"21.q_proj": 0.4684925973415375,
"21.k_proj": 0.40865975618362427,
"21.v_proj": 0.522392213344574,
"21.o_proj": 0.5564069747924805,
"21.gate_proj": 0.4738660454750061,
"21.up_proj": 0.49745243787765503,
"21.down_proj": 0.5537582039833069,
"22.q_proj": 0.5066864490509033,
"22.k_proj": 0.4351329803466797,
"22.v_proj": 0.5337282419204712,
"22.o_proj": 0.5765807628631592,
"22.gate_proj": 0.47878557443618774,
"22.up_proj": 0.4989929795265198,
"22.down_proj": 0.5528429746627808,
"23.q_proj": 0.49622949957847595,
"23.k_proj": 0.4328460097312927,
"23.v_proj": 0.5585790872573853,
"23.o_proj": 0.554423451423645,
"23.gate_proj": 0.48155754804611206,
"23.up_proj": 0.50072181224823,
"23.down_proj": 0.5464255213737488,
"24.q_proj": 0.5014330148696899,
"24.k_proj": 0.45270365476608276,
"24.v_proj": 0.569066047668457,
"24.o_proj": 0.5753594040870667,
"24.gate_proj": 0.4825673997402191,
"24.up_proj": 0.5023998618125916,
"24.down_proj": 0.5373293161392212,
"25.q_proj": 0.5077980756759644,
"25.k_proj": 0.4391773045063019,
"25.v_proj": 0.5462435483932495,
"25.o_proj": 0.5570780634880066,
"25.gate_proj": 0.4843604266643524,
"25.up_proj": 0.5035067796707153,
"25.down_proj": 0.5260589122772217,
"26.q_proj": 0.5026710033416748,
"26.k_proj": 0.4264448285102844,
"26.v_proj": 0.5314569473266602,
"26.o_proj": 0.537655234336853,
"26.gate_proj": 0.4852832555770874,
"26.up_proj": 0.5033767223358154,
"26.down_proj": 0.5124109387397766,
"27.q_proj": 0.5120652914047241,
"27.k_proj": 0.4639483690261841,
"27.v_proj": 0.5471333265304565,
"27.o_proj": 0.5303576588630676,
"27.gate_proj": 0.4886907935142517,
"27.up_proj": 0.5048775672912598,
"27.down_proj": 0.5119288563728333,
"28.q_proj": 0.4723275899887085,
"28.k_proj": 0.39252132177352905,
"28.v_proj": 0.5004826188087463,
"28.o_proj": 0.5055510997772217,
"28.gate_proj": 0.4891432821750641,
"28.up_proj": 0.5027062296867371,
"28.down_proj": 0.5162467956542969,
"29.q_proj": 0.45425698161125183,
"29.k_proj": 0.39296162128448486,
"29.v_proj": 0.529412031173706,
"29.o_proj": 0.4873389005661011,
"29.gate_proj": 0.4842919409275055,
"29.up_proj": 0.4974598288536072,
"29.down_proj": 0.5207532644271851,
"30.q_proj": 0.4333595335483551,
"30.k_proj": 0.35803964734077454,
"30.v_proj": 0.4631883203983307,
"30.o_proj": 0.47360387444496155,
"30.gate_proj": 0.4840942919254303,
"30.up_proj": 0.48931118845939636,
"30.down_proj": 0.5056915283203125,
"31.q_proj": 0.38473233580589294,
"31.k_proj": 0.3066442310810089,
"31.v_proj": 0.4090970456600189,
"31.o_proj": 0.3971359431743622,
"31.gate_proj": 0.4917009174823761,
"31.up_proj": 0.4686800241470337,
"31.down_proj": 0.4488861858844757
},
"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": 0.7003361628605769,
"mean_rel_err": 0.46043376145618303,
"max_rel_err": 0.5793784856796265
},
"factorisation": {
"max_rel_err": 0.00140141521114856,
"factor_params": 4887871488,
"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)"
}
}