svd-safety-l2_swift_remove30 / compression.json
Jeesup's picture
None arm at None budget
bda51eb 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": 0.3,
"internal_ratio_kept": 0.7,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/l2_swift_remove30/finetune_v/merge.pt",
"mean_rel_err": 0.45528166218926863,
"max_rel_err": 0.5735374689102173,
"param_fraction": 0.6997563653659327,
"ranks": {
"0.q_proj": 1697,
"0.k_proj": 1682,
"0.v_proj": 1605,
"0.o_proj": 1624,
"0.gate_proj": 2137,
"0.up_proj": 2137,
"0.down_proj": 2390,
"1.q_proj": 1568,
"1.k_proj": 1564,
"1.v_proj": 1509,
"1.o_proj": 1504,
"1.gate_proj": 2113,
"1.up_proj": 2112,
"1.down_proj": 2485,
"2.q_proj": 1431,
"2.k_proj": 1430,
"2.v_proj": 1407,
"2.o_proj": 1375,
"2.gate_proj": 2082,
"2.up_proj": 2082,
"2.down_proj": 1966,
"3.q_proj": 1440,
"3.k_proj": 1440,
"3.v_proj": 1430,
"3.o_proj": 1397,
"3.gate_proj": 2086,
"3.up_proj": 2086,
"3.down_proj": 2007,
"4.q_proj": 1472,
"4.k_proj": 1473,
"4.v_proj": 1461,
"4.o_proj": 1436,
"4.gate_proj": 2095,
"4.up_proj": 2094,
"4.down_proj": 2089,
"5.q_proj": 1442,
"5.k_proj": 1444,
"5.v_proj": 1433,
"5.o_proj": 1415,
"5.gate_proj": 2089,
"5.up_proj": 2089,
"5.down_proj": 2045,
"6.q_proj": 1466,
"6.k_proj": 1466,
"6.v_proj": 1462,
"6.o_proj": 1446,
"6.gate_proj": 2096,
"6.up_proj": 2094,
"6.down_proj": 2100,
"7.q_proj": 1434,
"7.k_proj": 1434,
"7.v_proj": 1431,
"7.o_proj": 1417,
"7.gate_proj": 2089,
"7.up_proj": 2088,
"7.down_proj": 2050,
"8.q_proj": 1446,
"8.k_proj": 1447,
"8.v_proj": 1445,
"8.o_proj": 1436,
"8.gate_proj": 2092,
"8.up_proj": 2091,
"8.down_proj": 2080,
"9.q_proj": 1423,
"9.k_proj": 1424,
"9.v_proj": 1423,
"9.o_proj": 1418,
"9.gate_proj": 2087,
"9.up_proj": 2086,
"9.down_proj": 2042,
"10.q_proj": 1411,
"10.k_proj": 1413,
"10.v_proj": 1412,
"10.o_proj": 1410,
"10.gate_proj": 2084,
"10.up_proj": 2084,
"10.down_proj": 2024,
"11.q_proj": 1411,
"11.k_proj": 1412,
"11.v_proj": 1415,
"11.o_proj": 1409,
"11.gate_proj": 2084,
"11.up_proj": 2084,
"11.down_proj": 2024,
"12.q_proj": 1404,
"12.k_proj": 1405,
"12.v_proj": 1408,
"12.o_proj": 1405,
"12.gate_proj": 2082,
"12.up_proj": 2082,
"12.down_proj": 2017,
"13.q_proj": 1408,
"13.k_proj": 1410,
"13.v_proj": 1413,
"13.o_proj": 1413,
"13.gate_proj": 2084,
"13.up_proj": 2084,
"13.down_proj": 2031,
"14.q_proj": 1404,
"14.k_proj": 1407,
"14.v_proj": 1407,
"14.o_proj": 1409,
"14.gate_proj": 2082,
"14.up_proj": 2082,
"14.down_proj": 2025,
"15.q_proj": 1405,
"15.k_proj": 1407,
"15.v_proj": 1409,
"15.o_proj": 1414,
"15.gate_proj": 2083,
"15.up_proj": 2083,
"15.down_proj": 2034,
"16.q_proj": 1405,
"16.k_proj": 1408,
"16.v_proj": 1410,
"16.o_proj": 1418,
"16.gate_proj": 2084,
"16.up_proj": 2084,
"16.down_proj": 2042,
"17.q_proj": 1402,
"17.k_proj": 1403,
"17.v_proj": 1408,
"17.o_proj": 1410,
"17.gate_proj": 2083,
"17.up_proj": 2083,
"17.down_proj": 2039,
"18.q_proj": 1400,
"18.k_proj": 1400,
"18.v_proj": 1407,
"18.o_proj": 1407,
"18.gate_proj": 2082,
"18.up_proj": 2082,
"18.down_proj": 2038,
"19.q_proj": 1392,
"19.k_proj": 1393,
"19.v_proj": 1400,
"19.o_proj": 1399,
"19.gate_proj": 2081,
"19.up_proj": 2081,
"19.down_proj": 2026,
"20.q_proj": 1391,
"20.k_proj": 1392,
"20.v_proj": 1398,
"20.o_proj": 1404,
"20.gate_proj": 2081,
"20.up_proj": 2081,
"20.down_proj": 2030,
"21.q_proj": 1387,
"21.k_proj": 1388,
"21.v_proj": 1397,
"21.o_proj": 1395,
"21.gate_proj": 2080,
"21.up_proj": 2080,
"21.down_proj": 2025,
"22.q_proj": 1388,
"22.k_proj": 1387,
"22.v_proj": 1397,
"22.o_proj": 1403,
"22.gate_proj": 2080,
"22.up_proj": 2080,
"22.down_proj": 2026,
"23.q_proj": 1385,
"23.k_proj": 1384,
"23.v_proj": 1398,
"23.o_proj": 1398,
"23.gate_proj": 2079,
"23.up_proj": 2079,
"23.down_proj": 2024,
"24.q_proj": 1383,
"24.k_proj": 1384,
"24.v_proj": 1396,
"24.o_proj": 1401,
"24.gate_proj": 2079,
"24.up_proj": 2079,
"24.down_proj": 2024,
"25.q_proj": 1384,
"25.k_proj": 1383,
"25.v_proj": 1399,
"25.o_proj": 1408,
"25.gate_proj": 2079,
"25.up_proj": 2080,
"25.down_proj": 2029,
"26.q_proj": 1385,
"26.k_proj": 1385,
"26.v_proj": 1399,
"26.o_proj": 1409,
"26.gate_proj": 2080,
"26.up_proj": 2080,
"26.down_proj": 2032,
"27.q_proj": 1386,
"27.k_proj": 1386,
"27.v_proj": 1401,
"27.o_proj": 1408,
"27.gate_proj": 2081,
"27.up_proj": 2081,
"27.down_proj": 2040,
"28.q_proj": 1387,
"28.k_proj": 1387,
"28.v_proj": 1404,
"28.o_proj": 1417,
"28.gate_proj": 2081,
"28.up_proj": 2082,
"28.down_proj": 2047,
"29.q_proj": 1388,
"29.k_proj": 1388,
"29.v_proj": 1404,
"29.o_proj": 1423,
"29.gate_proj": 2082,
"29.up_proj": 2083,
"29.down_proj": 2056,
"30.q_proj": 1403,
"30.k_proj": 1402,
"30.v_proj": 1422,
"30.o_proj": 1435,
"30.gate_proj": 2087,
"30.up_proj": 2088,
"30.down_proj": 2135,
"31.q_proj": 1728,
"31.k_proj": 1728,
"31.v_proj": 1746,
"31.o_proj": 1793,
"31.gate_proj": 2164,
"31.up_proj": 2167,
"31.down_proj": 2826
},
"rel_err": {
"0.q_proj": 0.13674421608448029,
"0.k_proj": 0.10850787907838821,
"0.v_proj": 0.2810197174549103,
"0.o_proj": 0.2883068323135376,
"0.gate_proj": 0.3820878565311432,
"0.up_proj": 0.3937278091907501,
"0.down_proj": 0.40569689869880676,
"1.q_proj": 0.11325579136610031,
"1.k_proj": 0.10081755369901657,
"1.v_proj": 0.4143904745578766,
"1.o_proj": 0.3389497399330139,
"1.gate_proj": 0.3668021559715271,
"1.up_proj": 0.4152345657348633,
"1.down_proj": 0.453666090965271,
"2.q_proj": 0.2440570592880249,
"2.k_proj": 0.2325412780046463,
"2.v_proj": 0.46705976128578186,
"2.o_proj": 0.5312212705612183,
"2.gate_proj": 0.4374466836452484,
"2.up_proj": 0.4691368341445923,
"2.down_proj": 0.5438665747642517,
"3.q_proj": 0.33946898579597473,
"3.k_proj": 0.33782610297203064,
"3.v_proj": 0.4778972566127777,
"3.o_proj": 0.5165881514549255,
"3.gate_proj": 0.48844465613365173,
"3.up_proj": 0.5085727572441101,
"3.down_proj": 0.5010958909988403,
"4.q_proj": 0.32332727313041687,
"4.k_proj": 0.32587742805480957,
"4.v_proj": 0.4671332538127899,
"4.o_proj": 0.5094566345214844,
"4.gate_proj": 0.48788368701934814,
"4.up_proj": 0.5130534768104553,
"4.down_proj": 0.48729878664016724,
"5.q_proj": 0.3404356837272644,
"5.k_proj": 0.3388126492500305,
"5.v_proj": 0.4766134023666382,
"5.o_proj": 0.5210713148117065,
"5.gate_proj": 0.4849763810634613,
"5.up_proj": 0.5152232646942139,
"5.down_proj": 0.48836052417755127,
"6.q_proj": 0.32434237003326416,
"6.k_proj": 0.33024048805236816,
"6.v_proj": 0.45296233892440796,
"6.o_proj": 0.4798653721809387,
"6.gate_proj": 0.47807908058166504,
"6.up_proj": 0.5142185091972351,
"6.down_proj": 0.4754852056503296,
"7.q_proj": 0.33590492606163025,
"7.k_proj": 0.34187355637550354,
"7.v_proj": 0.454049676656723,
"7.o_proj": 0.505239725112915,
"7.gate_proj": 0.4762929379940033,
"7.up_proj": 0.5152503848075867,
"7.down_proj": 0.4799094796180725,
"8.q_proj": 0.3426366448402405,
"8.k_proj": 0.3538832366466522,
"8.v_proj": 0.43721604347229004,
"8.o_proj": 0.4970349967479706,
"8.gate_proj": 0.47296127676963806,
"8.up_proj": 0.5100219249725342,
"8.down_proj": 0.46890032291412354,
"9.q_proj": 0.3625483214855194,
"9.k_proj": 0.3672429919242859,
"9.v_proj": 0.44050657749176025,
"9.o_proj": 0.5047832727432251,
"9.gate_proj": 0.47035032510757446,
"9.up_proj": 0.505872368812561,
"9.down_proj": 0.47603756189346313,
"10.q_proj": 0.3382200598716736,
"10.k_proj": 0.34823906421661377,
"10.v_proj": 0.4375532865524292,
"10.o_proj": 0.46764662861824036,
"10.gate_proj": 0.46595898270606995,
"10.up_proj": 0.5017201900482178,
"10.down_proj": 0.47638988494873047,
"11.q_proj": 0.32728126645088196,
"11.k_proj": 0.3394625782966614,
"11.v_proj": 0.4230077862739563,
"11.o_proj": 0.4579029679298401,
"11.gate_proj": 0.4656955301761627,
"11.up_proj": 0.5019961595535278,
"11.down_proj": 0.4826474189758301,
"12.q_proj": 0.36442241072654724,
"12.k_proj": 0.3636908531188965,
"12.v_proj": 0.4427216649055481,
"12.o_proj": 0.49870848655700684,
"12.gate_proj": 0.4701372981071472,
"12.up_proj": 0.5024581551551819,
"12.down_proj": 0.4839555323123932,
"13.q_proj": 0.3692917227745056,
"13.k_proj": 0.38481763005256653,
"13.v_proj": 0.46928614377975464,
"13.o_proj": 0.48683831095695496,
"13.gate_proj": 0.47251319885253906,
"13.up_proj": 0.5052496790885925,
"13.down_proj": 0.4925529658794403,
"14.q_proj": 0.35274773836135864,
"14.k_proj": 0.3569324016571045,
"14.v_proj": 0.46338796615600586,
"14.o_proj": 0.4851902723312378,
"14.gate_proj": 0.47242483496665955,
"14.up_proj": 0.5040963292121887,
"14.down_proj": 0.49548181891441345,
"15.q_proj": 0.37204164266586304,
"15.k_proj": 0.3740440607070923,
"15.v_proj": 0.4692009389400482,
"15.o_proj": 0.49231812357902527,
"15.gate_proj": 0.4763273298740387,
"15.up_proj": 0.5079472064971924,
"15.down_proj": 0.506531834602356,
"16.q_proj": 0.3697231411933899,
"16.k_proj": 0.37245628237724304,
"16.v_proj": 0.4960777759552002,
"16.o_proj": 0.5187640190124512,
"16.gate_proj": 0.4750789403915405,
"16.up_proj": 0.5048896670341492,
"16.down_proj": 0.516768753528595,
"17.q_proj": 0.40830114483833313,
"17.k_proj": 0.4063715934753418,
"17.v_proj": 0.5102226734161377,
"17.o_proj": 0.5319307446479797,
"17.gate_proj": 0.4774886965751648,
"17.up_proj": 0.505339503288269,
"17.down_proj": 0.5146914124488831,
"18.q_proj": 0.4319038391113281,
"18.k_proj": 0.43669185042381287,
"18.v_proj": 0.5166221261024475,
"18.o_proj": 0.5600407719612122,
"18.gate_proj": 0.48564109206199646,
"18.up_proj": 0.5114160180091858,
"18.down_proj": 0.5150787830352783,
"19.q_proj": 0.4182446599006653,
"19.k_proj": 0.4145455062389374,
"19.v_proj": 0.5035737156867981,
"19.o_proj": 0.5499191284179688,
"19.gate_proj": 0.4894777834415436,
"19.up_proj": 0.5121901035308838,
"19.down_proj": 0.5176525712013245,
"20.q_proj": 0.40922680497169495,
"20.k_proj": 0.41069889068603516,
"20.v_proj": 0.5174776911735535,
"20.o_proj": 0.5493062138557434,
"20.gate_proj": 0.48762205243110657,
"20.up_proj": 0.5089380741119385,
"20.down_proj": 0.5267716646194458,
"21.q_proj": 0.4238526523113251,
"21.k_proj": 0.4254571795463562,
"21.v_proj": 0.511239767074585,
"21.o_proj": 0.566460132598877,
"21.gate_proj": 0.48842716217041016,
"21.up_proj": 0.5083724856376648,
"21.down_proj": 0.5235127806663513,
"22.q_proj": 0.43453019857406616,
"22.k_proj": 0.438902884721756,
"22.v_proj": 0.5219279527664185,
"22.o_proj": 0.5610068440437317,
"22.gate_proj": 0.48826220631599426,
"22.up_proj": 0.5073332190513611,
"22.down_proj": 0.520425021648407,
"23.q_proj": 0.4653097093105316,
"23.k_proj": 0.46524491906166077,
"23.v_proj": 0.5198501944541931,
"23.o_proj": 0.5531998872756958,
"23.gate_proj": 0.49149203300476074,
"23.up_proj": 0.509039580821991,
"23.down_proj": 0.5131181478500366,
"24.q_proj": 0.4144188165664673,
"24.k_proj": 0.4183977246284485,
"24.v_proj": 0.5133265852928162,
"24.o_proj": 0.557678759098053,
"24.gate_proj": 0.4918333888053894,
"24.up_proj": 0.5094541907310486,
"24.down_proj": 0.5062175393104553,
"25.q_proj": 0.4548630118370056,
"25.k_proj": 0.46258294582366943,
"25.v_proj": 0.5203204154968262,
"25.o_proj": 0.548804521560669,
"25.gate_proj": 0.4945409893989563,
"25.up_proj": 0.5112118124961853,
"25.down_proj": 0.498342901468277,
"26.q_proj": 0.42560890316963196,
"26.k_proj": 0.4341040253639221,
"26.v_proj": 0.5099557042121887,
"26.o_proj": 0.5735374689102173,
"26.gate_proj": 0.49582305550575256,
"26.up_proj": 0.5132369995117188,
"26.down_proj": 0.4963560402393341,
"27.q_proj": 0.47353821992874146,
"27.k_proj": 0.4811285436153412,
"27.v_proj": 0.5287732481956482,
"27.o_proj": 0.5502170324325562,
"27.gate_proj": 0.49853190779685974,
"27.up_proj": 0.5145693421363831,
"27.down_proj": 0.49317970871925354,
"28.q_proj": 0.4549473822116852,
"28.k_proj": 0.46495798230171204,
"28.v_proj": 0.5245335102081299,
"28.o_proj": 0.5447710156440735,
"28.gate_proj": 0.4986145496368408,
"28.up_proj": 0.513562023639679,
"28.down_proj": 0.5004502534866333,
"29.q_proj": 0.42228034138679504,
"29.k_proj": 0.43027618527412415,
"29.v_proj": 0.5161402225494385,
"29.o_proj": 0.5447760224342346,
"29.gate_proj": 0.49749407172203064,
"29.up_proj": 0.5095394253730774,
"29.down_proj": 0.5059615969657898,
"30.q_proj": 0.45088398456573486,
"30.k_proj": 0.4586084485054016,
"30.v_proj": 0.5238063335418701,
"30.o_proj": 0.5494594573974609,
"30.gate_proj": 0.4911543130874634,
"30.up_proj": 0.49852705001831055,
"30.down_proj": 0.49798041582107544,
"31.q_proj": 0.3175491690635681,
"31.k_proj": 0.3234711289405823,
"31.v_proj": 0.41444289684295654,
"31.o_proj": 0.41846781969070435,
"31.gate_proj": 0.4670276939868927,
"31.up_proj": 0.47030943632125854,
"31.down_proj": 0.347734272480011
},
"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": 0.6997563653659327,
"mean_rel_err": 0.45483418411043075,
"max_rel_err": 0.5735054612159729
},
"factorisation": {
"max_rel_err": 0.0013921487843617797,
"factor_params": 4531625984,
"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)"
}
}