svd-safety-mis7_swift_remove20 / compression.json
Jeesup's picture
None arm at None budget
bdfdd18 verified
Raw History Blame Contribute Delete
15.3 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": 0.2,
"internal_ratio_kept": 0.8,
"from_pt": "/workspace/svd_safety/results/swift_ft/cells/mis7_swift_remove20/finetune_v/merge.pt",
"mean_rel_err": 0.40079765720292926,
"max_rel_err": 0.5248842835426331,
"param_fraction": 0.8004338191105769,
"ranks": {
"0.q_proj": 1843,
"0.k_proj": 776,
"0.v_proj": 706,
"0.o_proj": 1683,
"0.gate_proj": 2688,
"0.up_proj": 2688,
"0.down_proj": 2723,
"1.q_proj": 1851,
"1.k_proj": 789,
"1.v_proj": 743,
"1.o_proj": 1733,
"1.gate_proj": 2688,
"1.up_proj": 2688,
"1.down_proj": 3161,
"2.q_proj": 1508,
"2.k_proj": 598,
"2.v_proj": 587,
"2.o_proj": 1451,
"2.gate_proj": 2688,
"2.up_proj": 2688,
"2.down_proj": 2163,
"3.q_proj": 1504,
"3.k_proj": 595,
"3.v_proj": 581,
"3.o_proj": 1437,
"3.gate_proj": 2688,
"3.up_proj": 2688,
"3.down_proj": 2166,
"4.q_proj": 1501,
"4.k_proj": 594,
"4.v_proj": 586,
"4.o_proj": 1445,
"4.gate_proj": 2688,
"4.up_proj": 2688,
"4.down_proj": 2193,
"5.q_proj": 1501,
"5.k_proj": 595,
"5.v_proj": 586,
"5.o_proj": 1456,
"5.gate_proj": 2688,
"5.up_proj": 2688,
"5.down_proj": 2216,
"6.q_proj": 1503,
"6.k_proj": 596,
"6.v_proj": 588,
"6.o_proj": 1468,
"6.gate_proj": 2688,
"6.up_proj": 2688,
"6.down_proj": 2233,
"7.q_proj": 1501,
"7.k_proj": 597,
"7.v_proj": 589,
"7.o_proj": 1480,
"7.gate_proj": 2688,
"7.up_proj": 2688,
"7.down_proj": 2250,
"8.q_proj": 1497,
"8.k_proj": 593,
"8.v_proj": 589,
"8.o_proj": 1479,
"8.gate_proj": 2688,
"8.up_proj": 2688,
"8.down_proj": 2246,
"9.q_proj": 1494,
"9.k_proj": 592,
"9.v_proj": 590,
"9.o_proj": 1479,
"9.gate_proj": 2688,
"9.up_proj": 2688,
"9.down_proj": 2255,
"10.q_proj": 1496,
"10.k_proj": 593,
"10.v_proj": 590,
"10.o_proj": 1491,
"10.gate_proj": 2688,
"10.up_proj": 2688,
"10.down_proj": 2256,
"11.q_proj": 1494,
"11.k_proj": 592,
"11.v_proj": 592,
"11.o_proj": 1495,
"11.gate_proj": 2688,
"11.up_proj": 2688,
"11.down_proj": 2260,
"12.q_proj": 1503,
"12.k_proj": 598,
"12.v_proj": 597,
"12.o_proj": 1506,
"12.gate_proj": 2688,
"12.up_proj": 2688,
"12.down_proj": 2285,
"13.q_proj": 1498,
"13.k_proj": 596,
"13.v_proj": 595,
"13.o_proj": 1503,
"13.gate_proj": 2688,
"13.up_proj": 2688,
"13.down_proj": 2285,
"14.q_proj": 1510,
"14.k_proj": 601,
"14.v_proj": 604,
"14.o_proj": 1519,
"14.gate_proj": 2688,
"14.up_proj": 2688,
"14.down_proj": 2323,
"15.q_proj": 1527,
"15.k_proj": 612,
"15.v_proj": 615,
"15.o_proj": 1542,
"15.gate_proj": 2688,
"15.up_proj": 2688,
"15.down_proj": 2369,
"16.q_proj": 1534,
"16.k_proj": 615,
"16.v_proj": 618,
"16.o_proj": 1549,
"16.gate_proj": 2688,
"16.up_proj": 2688,
"16.down_proj": 2399,
"17.q_proj": 1565,
"17.k_proj": 631,
"17.v_proj": 635,
"17.o_proj": 1584,
"17.gate_proj": 2688,
"17.up_proj": 2688,
"17.down_proj": 2469,
"18.q_proj": 1565,
"18.k_proj": 630,
"18.v_proj": 636,
"18.o_proj": 1588,
"18.gate_proj": 2688,
"18.up_proj": 2688,
"18.down_proj": 2487,
"19.q_proj": 1544,
"19.k_proj": 620,
"19.v_proj": 626,
"19.o_proj": 1572,
"19.gate_proj": 2688,
"19.up_proj": 2688,
"19.down_proj": 2458,
"20.q_proj": 1507,
"20.k_proj": 599,
"20.v_proj": 606,
"20.o_proj": 1527,
"20.gate_proj": 2688,
"20.up_proj": 2688,
"20.down_proj": 2381,
"21.q_proj": 1497,
"21.k_proj": 593,
"21.v_proj": 601,
"21.o_proj": 1515,
"21.gate_proj": 2688,
"21.up_proj": 2688,
"21.down_proj": 2363,
"22.q_proj": 1482,
"22.k_proj": 583,
"22.v_proj": 594,
"22.o_proj": 1495,
"22.gate_proj": 2688,
"22.up_proj": 2688,
"22.down_proj": 2332,
"23.q_proj": 1483,
"23.k_proj": 584,
"23.v_proj": 593,
"23.o_proj": 1505,
"23.gate_proj": 2688,
"23.up_proj": 2688,
"23.down_proj": 2336,
"24.q_proj": 1482,
"24.k_proj": 584,
"24.v_proj": 594,
"24.o_proj": 1511,
"24.gate_proj": 2688,
"24.up_proj": 2688,
"24.down_proj": 2338,
"25.q_proj": 1486,
"25.k_proj": 586,
"25.v_proj": 597,
"25.o_proj": 1521,
"25.gate_proj": 2688,
"25.up_proj": 2688,
"25.down_proj": 2354,
"26.q_proj": 1493,
"26.k_proj": 589,
"26.v_proj": 602,
"26.o_proj": 1532,
"26.gate_proj": 2688,
"26.up_proj": 2688,
"26.down_proj": 2373,
"27.q_proj": 1498,
"27.k_proj": 592,
"27.v_proj": 606,
"27.o_proj": 1532,
"27.gate_proj": 2688,
"27.up_proj": 2688,
"27.down_proj": 2389,
"28.q_proj": 1518,
"28.k_proj": 603,
"28.v_proj": 618,
"28.o_proj": 1561,
"28.gate_proj": 2688,
"28.up_proj": 2688,
"28.down_proj": 2451,
"29.q_proj": 1530,
"29.k_proj": 609,
"29.v_proj": 626,
"29.o_proj": 1591,
"29.gate_proj": 2688,
"29.up_proj": 2688,
"29.down_proj": 2480,
"30.q_proj": 1553,
"30.k_proj": 621,
"30.v_proj": 642,
"30.o_proj": 1620,
"30.gate_proj": 2688,
"30.up_proj": 2688,
"30.down_proj": 2549,
"31.q_proj": 1684,
"31.k_proj": 692,
"31.v_proj": 716,
"31.o_proj": 1782,
"31.gate_proj": 2688,
"31.up_proj": 2688,
"31.down_proj": 2905
},
"rel_err": {
"0.q_proj": 0.08500257134437561,
"0.k_proj": 0.0716991052031517,
"0.v_proj": 0.22834858298301697,
"0.o_proj": 0.3019510805606842,
"0.gate_proj": 0.26146504282951355,
"0.up_proj": 0.2859381437301636,
"0.down_proj": 0.3900061547756195,
"1.q_proj": 0.13657206296920776,
"1.k_proj": 0.07040991634130478,
"1.v_proj": 0.34303975105285645,
"1.o_proj": 0.4063681960105896,
"1.gate_proj": 0.3133699595928192,
"1.up_proj": 0.33793723583221436,
"1.down_proj": 0.3351709246635437,
"2.q_proj": 0.35540199279785156,
"2.k_proj": 0.3083972930908203,
"2.v_proj": 0.4814577102661133,
"2.o_proj": 0.47539830207824707,
"2.gate_proj": 0.3777889311313629,
"2.up_proj": 0.39566245675086975,
"2.down_proj": 0.5091328620910645,
"3.q_proj": 0.3229297995567322,
"3.k_proj": 0.25059062242507935,
"3.v_proj": 0.5148245096206665,
"3.o_proj": 0.5196603536605835,
"3.gate_proj": 0.38781943917274475,
"3.up_proj": 0.4100722670555115,
"3.down_proj": 0.49922603368759155,
"4.q_proj": 0.3310073912143707,
"4.k_proj": 0.2755686044692993,
"4.v_proj": 0.4732750952243805,
"4.o_proj": 0.4666188955307007,
"4.gate_proj": 0.3899378180503845,
"4.up_proj": 0.4183727204799652,
"4.down_proj": 0.5054365992546082,
"5.q_proj": 0.3576917052268982,
"5.k_proj": 0.3083855211734772,
"5.v_proj": 0.5068833231925964,
"5.o_proj": 0.496389240026474,
"5.gate_proj": 0.3855798542499542,
"5.up_proj": 0.4201187789440155,
"5.down_proj": 0.4902642071247101,
"6.q_proj": 0.346874475479126,
"6.k_proj": 0.2829170227050781,
"6.v_proj": 0.4889771640300751,
"6.o_proj": 0.48441195487976074,
"6.gate_proj": 0.39520981907844543,
"6.up_proj": 0.4232458472251892,
"6.down_proj": 0.4810233414173126,
"7.q_proj": 0.33142852783203125,
"7.k_proj": 0.2655273973941803,
"7.v_proj": 0.4808761775493622,
"7.o_proj": 0.4556846618652344,
"7.gate_proj": 0.387683242559433,
"7.up_proj": 0.4178556799888611,
"7.down_proj": 0.4722205698490143,
"8.q_proj": 0.33769723773002625,
"8.k_proj": 0.2662927508354187,
"8.v_proj": 0.4425075054168701,
"8.o_proj": 0.44497787952423096,
"8.gate_proj": 0.3896368443965912,
"8.up_proj": 0.41572481393814087,
"8.down_proj": 0.4673706591129303,
"9.q_proj": 0.3340452313423157,
"9.k_proj": 0.2718675434589386,
"9.v_proj": 0.4268815815448761,
"9.o_proj": 0.44764506816864014,
"9.gate_proj": 0.38748565316200256,
"9.up_proj": 0.4152057468891144,
"9.down_proj": 0.4528351426124573,
"10.q_proj": 0.3335397243499756,
"10.k_proj": 0.2776794731616974,
"10.v_proj": 0.4279860556125641,
"10.o_proj": 0.44602537155151367,
"10.gate_proj": 0.38787901401519775,
"10.up_proj": 0.4141860604286194,
"10.down_proj": 0.46012479066848755,
"11.q_proj": 0.3297421634197235,
"11.k_proj": 0.26538947224617004,
"11.v_proj": 0.3986572325229645,
"11.o_proj": 0.4476287066936493,
"11.gate_proj": 0.38680461049079895,
"11.up_proj": 0.41310203075408936,
"11.down_proj": 0.45812714099884033,
"12.q_proj": 0.319181889295578,
"12.k_proj": 0.24617546796798706,
"12.v_proj": 0.4383234977722168,
"12.o_proj": 0.42985114455223083,
"12.gate_proj": 0.3842836320400238,
"12.up_proj": 0.4086284637451172,
"12.down_proj": 0.45287853479385376,
"13.q_proj": 0.342117041349411,
"13.k_proj": 0.28727155923843384,
"13.v_proj": 0.4403918981552124,
"13.o_proj": 0.43300414085388184,
"13.gate_proj": 0.3891292214393616,
"13.up_proj": 0.4081304669380188,
"13.down_proj": 0.4504336416721344,
"14.q_proj": 0.31687843799591064,
"14.k_proj": 0.26946091651916504,
"14.v_proj": 0.44405218958854675,
"14.o_proj": 0.43850401043891907,
"14.gate_proj": 0.38567572832107544,
"14.up_proj": 0.4120667278766632,
"14.down_proj": 0.4714415967464447,
"15.q_proj": 0.31565016508102417,
"15.k_proj": 0.2453945130109787,
"15.v_proj": 0.4641188681125641,
"15.o_proj": 0.4399622678756714,
"15.gate_proj": 0.38257908821105957,
"15.up_proj": 0.408530592918396,
"15.down_proj": 0.4638931155204773,
"16.q_proj": 0.32283511757850647,
"16.k_proj": 0.2652204930782318,
"16.v_proj": 0.464958131313324,
"16.o_proj": 0.42660167813301086,
"16.gate_proj": 0.3851650655269623,
"16.up_proj": 0.41526761651039124,
"16.down_proj": 0.46140268445014954,
"17.q_proj": 0.3427833616733551,
"17.k_proj": 0.2890065610408783,
"17.v_proj": 0.4380885064601898,
"17.o_proj": 0.44451603293418884,
"17.gate_proj": 0.3959171772003174,
"17.up_proj": 0.4251500368118286,
"17.down_proj": 0.4505082368850708,
"18.q_proj": 0.3227398097515106,
"18.k_proj": 0.2645529508590698,
"18.v_proj": 0.486581027507782,
"18.o_proj": 0.4234537184238434,
"18.gate_proj": 0.39415696263313293,
"18.up_proj": 0.42284369468688965,
"18.down_proj": 0.4652669429779053,
"19.q_proj": 0.36690786480903625,
"19.k_proj": 0.27805012464523315,
"19.v_proj": 0.5099502205848694,
"19.o_proj": 0.4442979693412781,
"19.gate_proj": 0.39896416664123535,
"19.up_proj": 0.42378658056259155,
"19.down_proj": 0.48121050000190735,
"20.q_proj": 0.38949716091156006,
"20.k_proj": 0.32147786021232605,
"20.v_proj": 0.4748760461807251,
"20.o_proj": 0.46483227610588074,
"20.gate_proj": 0.4035407602787018,
"20.up_proj": 0.423808217048645,
"20.down_proj": 0.49269551038742065,
"21.q_proj": 0.41730770468711853,
"21.k_proj": 0.3455001711845398,
"21.v_proj": 0.45047393441200256,
"21.o_proj": 0.5031458139419556,
"21.gate_proj": 0.4041469395160675,
"21.up_proj": 0.4253368079662323,
"21.down_proj": 0.4889674484729767,
"22.q_proj": 0.4534510374069214,
"22.k_proj": 0.37348392605781555,
"22.v_proj": 0.4685239791870117,
"22.o_proj": 0.5248842835426331,
"22.gate_proj": 0.4087425768375397,
"22.up_proj": 0.42668619751930237,
"22.down_proj": 0.4892253875732422,
"23.q_proj": 0.4440714418888092,
"23.k_proj": 0.3682272434234619,
"23.v_proj": 0.49346232414245605,
"23.o_proj": 0.502592921257019,
"23.gate_proj": 0.4109232425689697,
"23.up_proj": 0.42807736992836,
"23.down_proj": 0.48296502232551575,
"24.q_proj": 0.4487209916114807,
"24.k_proj": 0.38935786485671997,
"24.v_proj": 0.5036438703536987,
"24.o_proj": 0.5222789645195007,
"24.gate_proj": 0.4119880795478821,
"24.up_proj": 0.42954087257385254,
"24.down_proj": 0.47408875823020935,
"25.q_proj": 0.4551296532154083,
"25.k_proj": 0.37871479988098145,
"25.v_proj": 0.4806521534919739,
"25.o_proj": 0.5046061873435974,
"25.gate_proj": 0.4138759672641754,
"25.up_proj": 0.4307102560997009,
"25.down_proj": 0.463013231754303,
"26.q_proj": 0.44990450143814087,
"26.k_proj": 0.3649565577507019,
"26.v_proj": 0.46789830923080444,
"26.o_proj": 0.4847952425479889,
"26.gate_proj": 0.414262056350708,
"26.up_proj": 0.43036970496177673,
"26.down_proj": 0.4495799243450165,
"27.q_proj": 0.4583161473274231,
"27.k_proj": 0.39734572172164917,
"27.v_proj": 0.4797323942184448,
"27.o_proj": 0.47943976521492004,
"27.gate_proj": 0.41736742854118347,
"27.up_proj": 0.4318309724330902,
"27.down_proj": 0.44924622774124146,
"28.q_proj": 0.41977617144584656,
"28.k_proj": 0.3314531147480011,
"28.v_proj": 0.4329354465007782,
"28.o_proj": 0.4519560635089874,
"28.gate_proj": 0.41815441846847534,
"28.up_proj": 0.4296439588069916,
"28.down_proj": 0.4507562816143036,
"29.q_proj": 0.4013705253601074,
"29.k_proj": 0.3302054703235626,
"29.v_proj": 0.45994922518730164,
"29.o_proj": 0.4338970482349396,
"29.gate_proj": 0.4139324426651001,
"29.up_proj": 0.4252754747867584,
"29.down_proj": 0.4538421928882599,
"30.q_proj": 0.3819069266319275,
"30.k_proj": 0.300334095954895,
"30.v_proj": 0.3898763358592987,
"30.o_proj": 0.4189046621322632,
"30.gate_proj": 0.4136664569377899,
"30.up_proj": 0.4180448055267334,
"30.down_proj": 0.43752264976501465,
"31.q_proj": 0.3336416184902191,
"31.k_proj": 0.2430993765592575,
"31.v_proj": 0.32917752861976624,
"31.o_proj": 0.3434261381626129,
"31.gate_proj": 0.4216300845146179,
"31.up_proj": 0.40061134099960327,
"31.down_proj": 0.3710506558418274
},
"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": 0.8004338191105769,
"mean_rel_err": 0.3994709092990628,
"max_rel_err": 0.5239740610122681
},
"factorisation": {
"max_rel_err": 0.0013197725638747215,
"factor_params": 5586485248,
"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)"
}
}