svd-safety-l2_swift_remove20 / compression.json
Jeesup's picture
None arm at None budget
abfa9b8 verified
Raw History Blame Contribute Delete
15.3 kB
{
"source": "/workspace/svd_safety/assets/models/Llama-2-7b-chat-hf",
"upstream": "hiahei/Swift-SVD (compression) + AIoT-MLSys-Lab/SVD-LLM utils/LoRA.py (recovery)",
"stage": "finetuned",
"ratio_removed": 0.2,
"internal_ratio_kept": 0.8,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/l2_swift_remove20/finetune_v/merge.pt",
"mean_rel_err": 0.39835471925990923,
"max_rel_err": 0.5163237452507019,
"param_fraction": 0.7998957658678757,
"ranks": {
"0.q_proj": 1939,
"0.k_proj": 1922,
"0.v_proj": 1834,
"0.o_proj": 1856,
"0.gate_proj": 2443,
"0.up_proj": 2444,
"0.down_proj": 2732,
"1.q_proj": 1792,
"1.k_proj": 1788,
"1.v_proj": 1724,
"1.o_proj": 1719,
"1.gate_proj": 2415,
"1.up_proj": 2414,
"1.down_proj": 2841,
"2.q_proj": 1635,
"2.k_proj": 1635,
"2.v_proj": 1609,
"2.o_proj": 1572,
"2.gate_proj": 2380,
"2.up_proj": 2380,
"2.down_proj": 2247,
"3.q_proj": 1646,
"3.k_proj": 1646,
"3.v_proj": 1634,
"3.o_proj": 1597,
"3.gate_proj": 2385,
"3.up_proj": 2385,
"3.down_proj": 2294,
"4.q_proj": 1682,
"4.k_proj": 1684,
"4.v_proj": 1670,
"4.o_proj": 1641,
"4.gate_proj": 2395,
"4.up_proj": 2394,
"4.down_proj": 2388,
"5.q_proj": 1648,
"5.k_proj": 1650,
"5.v_proj": 1638,
"5.o_proj": 1617,
"5.gate_proj": 2388,
"5.up_proj": 2387,
"5.down_proj": 2338,
"6.q_proj": 1675,
"6.k_proj": 1675,
"6.v_proj": 1671,
"6.o_proj": 1653,
"6.gate_proj": 2395,
"6.up_proj": 2394,
"6.down_proj": 2401,
"7.q_proj": 1639,
"7.k_proj": 1639,
"7.v_proj": 1636,
"7.o_proj": 1619,
"7.gate_proj": 2388,
"7.up_proj": 2386,
"7.down_proj": 2344,
"8.q_proj": 1653,
"8.k_proj": 1654,
"8.v_proj": 1652,
"8.o_proj": 1641,
"8.gate_proj": 2391,
"8.up_proj": 2390,
"8.down_proj": 2377,
"9.q_proj": 1626,
"9.k_proj": 1628,
"9.v_proj": 1627,
"9.o_proj": 1621,
"9.gate_proj": 2386,
"9.up_proj": 2385,
"9.down_proj": 2335,
"10.q_proj": 1613,
"10.k_proj": 1615,
"10.v_proj": 1614,
"10.o_proj": 1613,
"10.gate_proj": 2383,
"10.up_proj": 2382,
"10.down_proj": 2314,
"11.q_proj": 1612,
"11.k_proj": 1613,
"11.v_proj": 1617,
"11.o_proj": 1610,
"11.gate_proj": 2383,
"11.up_proj": 2382,
"11.down_proj": 2314,
"12.q_proj": 1605,
"12.k_proj": 1607,
"12.v_proj": 1610,
"12.o_proj": 1606,
"12.gate_proj": 2380,
"12.up_proj": 2380,
"12.down_proj": 2306,
"13.q_proj": 1611,
"13.k_proj": 1612,
"13.v_proj": 1615,
"13.o_proj": 1616,
"13.gate_proj": 2382,
"13.up_proj": 2382,
"13.down_proj": 2321,
"14.q_proj": 1605,
"14.k_proj": 1607,
"14.v_proj": 1609,
"14.o_proj": 1610,
"14.gate_proj": 2380,
"14.up_proj": 2380,
"14.down_proj": 2315,
"15.q_proj": 1607,
"15.k_proj": 1608,
"15.v_proj": 1612,
"15.o_proj": 1616,
"15.gate_proj": 2382,
"15.up_proj": 2382,
"15.down_proj": 2324,
"16.q_proj": 1607,
"16.k_proj": 1610,
"16.v_proj": 1612,
"16.o_proj": 1621,
"16.gate_proj": 2382,
"16.up_proj": 2382,
"16.down_proj": 2334,
"17.q_proj": 1603,
"17.k_proj": 1604,
"17.v_proj": 1609,
"17.o_proj": 1612,
"17.gate_proj": 2381,
"17.up_proj": 2381,
"17.down_proj": 2330,
"18.q_proj": 1600,
"18.k_proj": 1601,
"18.v_proj": 1608,
"18.o_proj": 1608,
"18.gate_proj": 2380,
"18.up_proj": 2380,
"18.down_proj": 2329,
"19.q_proj": 1591,
"19.k_proj": 1592,
"19.v_proj": 1600,
"19.o_proj": 1600,
"19.gate_proj": 2378,
"19.up_proj": 2378,
"19.down_proj": 2316,
"20.q_proj": 1590,
"20.k_proj": 1591,
"20.v_proj": 1598,
"20.o_proj": 1605,
"20.gate_proj": 2378,
"20.up_proj": 2378,
"20.down_proj": 2320,
"21.q_proj": 1586,
"21.k_proj": 1586,
"21.v_proj": 1597,
"21.o_proj": 1595,
"21.gate_proj": 2378,
"21.up_proj": 2378,
"21.down_proj": 2315,
"22.q_proj": 1586,
"22.k_proj": 1586,
"22.v_proj": 1597,
"22.o_proj": 1603,
"22.gate_proj": 2377,
"22.up_proj": 2378,
"22.down_proj": 2317,
"23.q_proj": 1583,
"23.k_proj": 1582,
"23.v_proj": 1598,
"23.o_proj": 1598,
"23.gate_proj": 2377,
"23.up_proj": 2377,
"23.down_proj": 2314,
"24.q_proj": 1581,
"24.k_proj": 1582,
"24.v_proj": 1595,
"24.o_proj": 1602,
"24.gate_proj": 2377,
"24.up_proj": 2377,
"24.down_proj": 2314,
"25.q_proj": 1582,
"25.k_proj": 1581,
"25.v_proj": 1599,
"25.o_proj": 1610,
"25.gate_proj": 2377,
"25.up_proj": 2377,
"25.down_proj": 2320,
"26.q_proj": 1583,
"26.k_proj": 1583,
"26.v_proj": 1599,
"26.o_proj": 1611,
"26.gate_proj": 2378,
"26.up_proj": 2378,
"26.down_proj": 2323,
"27.q_proj": 1585,
"27.k_proj": 1584,
"27.v_proj": 1602,
"27.o_proj": 1609,
"27.gate_proj": 2378,
"27.up_proj": 2379,
"27.down_proj": 2331,
"28.q_proj": 1586,
"28.k_proj": 1586,
"28.v_proj": 1605,
"28.o_proj": 1619,
"28.gate_proj": 2379,
"28.up_proj": 2380,
"28.down_proj": 2340,
"29.q_proj": 1587,
"29.k_proj": 1587,
"29.v_proj": 1605,
"29.o_proj": 1627,
"29.gate_proj": 2380,
"29.up_proj": 2382,
"29.down_proj": 2350,
"30.q_proj": 1604,
"30.k_proj": 1603,
"30.v_proj": 1625,
"30.o_proj": 1640,
"30.gate_proj": 2386,
"30.up_proj": 2387,
"30.down_proj": 2441,
"31.q_proj": 1974,
"31.k_proj": 1975,
"31.v_proj": 1995,
"31.o_proj": 2049,
"31.gate_proj": 2474,
"31.up_proj": 2477,
"31.down_proj": 3231
},
"rel_err": {
"0.q_proj": 0.13252992928028107,
"0.k_proj": 0.10504010319709778,
"0.v_proj": 0.23756639659404755,
"0.o_proj": 0.2499486804008484,
"0.gate_proj": 0.32737889885902405,
"0.up_proj": 0.336822509765625,
"0.down_proj": 0.33574700355529785,
"1.q_proj": 0.09638579189777374,
"1.k_proj": 0.0869532972574234,
"1.v_proj": 0.3620476722717285,
"1.o_proj": 0.2894635498523712,
"1.gate_proj": 0.31223469972610474,
"1.up_proj": 0.3540322184562683,
"1.down_proj": 0.3827917277812958,
"2.q_proj": 0.20203469693660736,
"2.k_proj": 0.19166691601276398,
"2.v_proj": 0.4135085940361023,
"2.o_proj": 0.4775710105895996,
"2.gate_proj": 0.38029783964157104,
"2.up_proj": 0.40832197666168213,
"2.down_proj": 0.4817647933959961,
"3.q_proj": 0.28911951184272766,
"3.k_proj": 0.2867107093334198,
"3.v_proj": 0.42236825823783875,
"3.o_proj": 0.4629066586494446,
"3.gate_proj": 0.4270183742046356,
"3.up_proj": 0.4446209967136383,
"3.down_proj": 0.43980148434638977,
"4.q_proj": 0.2749663293361664,
"4.k_proj": 0.27662697434425354,
"4.v_proj": 0.410868376493454,
"4.o_proj": 0.45410028100013733,
"4.gate_proj": 0.4263415038585663,
"4.up_proj": 0.449140727519989,
"4.down_proj": 0.4236817955970764,
"5.q_proj": 0.2918066382408142,
"5.k_proj": 0.2901315689086914,
"5.v_proj": 0.4196918308734894,
"5.o_proj": 0.46734270453453064,
"5.gate_proj": 0.42409467697143555,
"5.up_proj": 0.4514884650707245,
"5.down_proj": 0.42588138580322266,
"6.q_proj": 0.27811768651008606,
"6.k_proj": 0.2831922173500061,
"6.v_proj": 0.39679214358329773,
"6.o_proj": 0.42545533180236816,
"6.gate_proj": 0.41754117608070374,
"6.up_proj": 0.44990018010139465,
"6.down_proj": 0.4116297662258148,
"7.q_proj": 0.2889671325683594,
"7.k_proj": 0.2941109836101532,
"7.v_proj": 0.3983648121356964,
"7.o_proj": 0.45035600662231445,
"7.gate_proj": 0.4161233603954315,
"7.up_proj": 0.4515121281147003,
"7.down_proj": 0.41752707958221436,
"8.q_proj": 0.2945907711982727,
"8.k_proj": 0.30408018827438354,
"8.v_proj": 0.38267946243286133,
"8.o_proj": 0.44145527482032776,
"8.gate_proj": 0.41320326924324036,
"8.up_proj": 0.4467734396457672,
"8.down_proj": 0.406986266374588,
"9.q_proj": 0.3136454224586487,
"9.k_proj": 0.31721559166908264,
"9.v_proj": 0.3859423100948334,
"9.o_proj": 0.4510596990585327,
"9.gate_proj": 0.4109175205230713,
"9.up_proj": 0.4431936740875244,
"9.down_proj": 0.4149797558784485,
"10.q_proj": 0.29216355085372925,
"10.k_proj": 0.3005751371383667,
"10.v_proj": 0.3847027122974396,
"10.o_proj": 0.4111638367176056,
"10.gate_proj": 0.4067463278770447,
"10.up_proj": 0.43945708870887756,
"10.down_proj": 0.4159848392009735,
"11.q_proj": 0.2826056480407715,
"11.k_proj": 0.2936619520187378,
"11.v_proj": 0.3708671033382416,
"11.o_proj": 0.40357112884521484,
"11.gate_proj": 0.40603089332580566,
"11.up_proj": 0.4393159747123718,
"11.down_proj": 0.42167046666145325,
"12.q_proj": 0.31589677929878235,
"12.k_proj": 0.3152616322040558,
"12.v_proj": 0.38857823610305786,
"12.o_proj": 0.4415610432624817,
"12.gate_proj": 0.4102623462677002,
"12.up_proj": 0.43980276584625244,
"12.down_proj": 0.4229101836681366,
"13.q_proj": 0.320039302110672,
"13.k_proj": 0.33416274189949036,
"13.v_proj": 0.41291582584381104,
"13.o_proj": 0.4305266737937927,
"13.gate_proj": 0.4123358428478241,
"13.up_proj": 0.4423108994960785,
"13.down_proj": 0.43072378635406494,
"14.q_proj": 0.30624768137931824,
"14.k_proj": 0.31049633026123047,
"14.v_proj": 0.4073112905025482,
"14.o_proj": 0.4292844831943512,
"14.gate_proj": 0.41197630763053894,
"14.up_proj": 0.44101130962371826,
"14.down_proj": 0.4332798719406128,
"15.q_proj": 0.32311615347862244,
"15.k_proj": 0.325520783662796,
"15.v_proj": 0.41313716769218445,
"15.o_proj": 0.43440884351730347,
"15.gate_proj": 0.415483683347702,
"15.up_proj": 0.44479092955589294,
"15.down_proj": 0.44361311197280884,
"16.q_proj": 0.32134386897087097,
"16.k_proj": 0.32373708486557007,
"16.v_proj": 0.4383922517299652,
"16.o_proj": 0.460711807012558,
"16.gate_proj": 0.41429081559181213,
"16.up_proj": 0.4417015612125397,
"16.down_proj": 0.4521862864494324,
"17.q_proj": 0.35595712065696716,
"17.k_proj": 0.3545474112033844,
"17.v_proj": 0.4520525336265564,
"17.o_proj": 0.4732700288295746,
"17.gate_proj": 0.41633883118629456,
"17.up_proj": 0.4417755901813507,
"17.down_proj": 0.4500233232975006,
"18.q_proj": 0.37810665369033813,
"18.k_proj": 0.3822422921657562,
"18.v_proj": 0.45829951763153076,
"18.o_proj": 0.5030460357666016,
"18.gate_proj": 0.4236604869365692,
"18.up_proj": 0.4474699795246124,
"18.down_proj": 0.450743168592453,
"19.q_proj": 0.3656962513923645,
"19.k_proj": 0.3624645471572876,
"19.v_proj": 0.4467334449291229,
"19.o_proj": 0.4933334290981293,
"19.gate_proj": 0.42747756838798523,
"19.up_proj": 0.44855284690856934,
"19.down_proj": 0.4535249173641205,
"20.q_proj": 0.3567904829978943,
"20.k_proj": 0.35822567343711853,
"20.v_proj": 0.45949605107307434,
"20.o_proj": 0.4911755919456482,
"20.gate_proj": 0.42556217312812805,
"20.up_proj": 0.44537943601608276,
"20.down_proj": 0.4616205394268036,
"21.q_proj": 0.3698049783706665,
"21.k_proj": 0.371379554271698,
"21.v_proj": 0.45389649271965027,
"21.o_proj": 0.5087231397628784,
"21.gate_proj": 0.42624449729919434,
"21.up_proj": 0.44437286257743835,
"21.down_proj": 0.45840954780578613,
"22.q_proj": 0.37954530119895935,
"22.k_proj": 0.38333773612976074,
"22.v_proj": 0.463877409696579,
"22.o_proj": 0.5026487112045288,
"22.gate_proj": 0.4261413812637329,
"22.up_proj": 0.443158358335495,
"22.down_proj": 0.4550263583660126,
"23.q_proj": 0.4084702432155609,
"23.k_proj": 0.40841731429100037,
"23.v_proj": 0.4616864025592804,
"23.o_proj": 0.4950319528579712,
"23.gate_proj": 0.4287131428718567,
"23.up_proj": 0.444684237241745,
"23.down_proj": 0.44846802949905396,
"24.q_proj": 0.36004146933555603,
"24.k_proj": 0.363835871219635,
"24.v_proj": 0.45619717240333557,
"24.o_proj": 0.4993390142917633,
"24.gate_proj": 0.4288971722126007,
"24.up_proj": 0.4449731707572937,
"24.down_proj": 0.44221973419189453,
"25.q_proj": 0.39867693185806274,
"25.k_proj": 0.40591371059417725,
"25.v_proj": 0.4620579481124878,
"25.o_proj": 0.491049200296402,
"25.gate_proj": 0.4311861991882324,
"25.up_proj": 0.44645044207572937,
"25.down_proj": 0.43493691086769104,
"26.q_proj": 0.37144312262535095,
"26.k_proj": 0.37911906838417053,
"26.v_proj": 0.45244038105010986,
"26.o_proj": 0.5163237452507019,
"26.gate_proj": 0.4322204887866974,
"26.up_proj": 0.44827595353126526,
"26.down_proj": 0.4333139657974243,
"27.q_proj": 0.4169563055038452,
"27.k_proj": 0.4245080351829529,
"27.v_proj": 0.46958377957344055,
"27.o_proj": 0.49201327562332153,
"27.gate_proj": 0.4350823760032654,
"27.up_proj": 0.44976213574409485,
"27.down_proj": 0.43056654930114746,
"28.q_proj": 0.399061918258667,
"28.k_proj": 0.4083995521068573,
"28.v_proj": 0.46559157967567444,
"28.o_proj": 0.4871852695941925,
"28.gate_proj": 0.4352478086948395,
"28.up_proj": 0.4488693177700043,
"28.down_proj": 0.4368627965450287,
"29.q_proj": 0.36882483959198,
"29.k_proj": 0.37630507349967957,
"29.v_proj": 0.4574759304523468,
"29.o_proj": 0.48539915680885315,
"29.gate_proj": 0.43431898951530457,
"29.up_proj": 0.4452069103717804,
"29.down_proj": 0.4417956471443176,
"30.q_proj": 0.3955976068973541,
"30.k_proj": 0.4027300477027893,
"30.v_proj": 0.4643517732620239,
"30.o_proj": 0.4904117286205292,
"30.gate_proj": 0.42837339639663696,
"30.up_proj": 0.4352115988731384,
"30.down_proj": 0.430807888507843,
"31.q_proj": 0.2638401985168457,
"31.k_proj": 0.26833730936050415,
"31.v_proj": 0.34875962138175964,
"31.o_proj": 0.35200828313827515,
"31.gate_proj": 0.40333712100982666,
"31.up_proj": 0.4065763056278229,
"31.down_proj": 0.26060518622398376
},
"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": null,
"param_fraction": 0.7993212685066183,
"mean_rel_err": 0.39799320328581544,
"max_rel_err": 0.516345202922821
},
"factorisation": {
"max_rel_err": 0.0013380160089582205,
"factor_params": 5180129280,
"dense_params": 6476005376
},
"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)"
}
}