svd-safety-l2_basis_remove20 / compression.json
Jeesup's picture
None arm at None budget
c161224 verified
Raw History Blame Contribute Delete
15.9 kB
{
"source": "/workspace/svd_safety/assets/models/Llama-2-7b-chat-hf",
"upstream": "TUDa-HWAI/Basis_Sharing",
"step": "basis_sharing",
"ratio_removed": 0.2,
"group_size": 2,
"nsamples": 256,
"seqlen": 2048,
"seed": 42,
"share_part": [
"v",
"k",
"q",
"up",
"gate"
],
"private_part": [
"down",
"o"
],
"qkv_bias_restored": 0,
"factored": "/workspace/svd_safety/results/basis_sharing/cells/l2_basis_remove20/truncated/factored.pt",
"param_fraction": 0.7998198672279793,
"ranks": {
"0.q_proj": 2184,
"0.k_proj": 2184,
"0.v_proj": 2184,
"0.o_proj": 1638,
"0.gate_proj": 2762,
"0.up_proj": 2762,
"0.down_proj": 2388,
"1.q_proj": 2184,
"1.k_proj": 2184,
"1.v_proj": 2184,
"1.o_proj": 1638,
"1.gate_proj": 2762,
"1.up_proj": 2762,
"1.down_proj": 2388,
"2.q_proj": 2184,
"2.k_proj": 2184,
"2.v_proj": 2184,
"2.o_proj": 1638,
"2.gate_proj": 2762,
"2.up_proj": 2762,
"2.down_proj": 2388,
"3.q_proj": 2184,
"3.k_proj": 2184,
"3.v_proj": 2184,
"3.o_proj": 1638,
"3.gate_proj": 2762,
"3.up_proj": 2762,
"3.down_proj": 2388,
"4.q_proj": 2184,
"4.k_proj": 2184,
"4.v_proj": 2184,
"4.o_proj": 1638,
"4.gate_proj": 2762,
"4.up_proj": 2762,
"4.down_proj": 2388,
"5.q_proj": 2184,
"5.k_proj": 2184,
"5.v_proj": 2184,
"5.o_proj": 1638,
"5.gate_proj": 2762,
"5.up_proj": 2762,
"5.down_proj": 2388,
"6.q_proj": 2184,
"6.k_proj": 2184,
"6.v_proj": 2184,
"6.o_proj": 1638,
"6.gate_proj": 2762,
"6.up_proj": 2762,
"6.down_proj": 2388,
"7.q_proj": 2184,
"7.k_proj": 2184,
"7.v_proj": 2184,
"7.o_proj": 1638,
"7.gate_proj": 2762,
"7.up_proj": 2762,
"7.down_proj": 2388,
"8.q_proj": 2184,
"8.k_proj": 2184,
"8.v_proj": 2184,
"8.o_proj": 1638,
"8.gate_proj": 2762,
"8.up_proj": 2762,
"8.down_proj": 2388,
"9.q_proj": 2184,
"9.k_proj": 2184,
"9.v_proj": 2184,
"9.o_proj": 1638,
"9.gate_proj": 2762,
"9.up_proj": 2762,
"9.down_proj": 2388,
"10.q_proj": 2184,
"10.k_proj": 2184,
"10.v_proj": 2184,
"10.o_proj": 1638,
"10.gate_proj": 2762,
"10.up_proj": 2762,
"10.down_proj": 2388,
"11.q_proj": 2184,
"11.k_proj": 2184,
"11.v_proj": 2184,
"11.o_proj": 1638,
"11.gate_proj": 2762,
"11.up_proj": 2762,
"11.down_proj": 2388,
"12.q_proj": 2184,
"12.k_proj": 2184,
"12.v_proj": 2184,
"12.o_proj": 1638,
"12.gate_proj": 2762,
"12.up_proj": 2762,
"12.down_proj": 2388,
"13.q_proj": 2184,
"13.k_proj": 2184,
"13.v_proj": 2184,
"13.o_proj": 1638,
"13.gate_proj": 2762,
"13.up_proj": 2762,
"13.down_proj": 2388,
"14.q_proj": 2184,
"14.k_proj": 2184,
"14.v_proj": 2184,
"14.o_proj": 1638,
"14.gate_proj": 2762,
"14.up_proj": 2762,
"14.down_proj": 2388,
"15.q_proj": 2184,
"15.k_proj": 2184,
"15.v_proj": 2184,
"15.o_proj": 1638,
"15.gate_proj": 2762,
"15.up_proj": 2762,
"15.down_proj": 2388,
"16.q_proj": 2184,
"16.k_proj": 2184,
"16.v_proj": 2184,
"16.o_proj": 1638,
"16.gate_proj": 2762,
"16.up_proj": 2762,
"16.down_proj": 2388,
"17.q_proj": 2184,
"17.k_proj": 2184,
"17.v_proj": 2184,
"17.o_proj": 1638,
"17.gate_proj": 2762,
"17.up_proj": 2762,
"17.down_proj": 2388,
"18.q_proj": 2184,
"18.k_proj": 2184,
"18.v_proj": 2184,
"18.o_proj": 1638,
"18.gate_proj": 2762,
"18.up_proj": 2762,
"18.down_proj": 2388,
"19.q_proj": 2184,
"19.k_proj": 2184,
"19.v_proj": 2184,
"19.o_proj": 1638,
"19.gate_proj": 2762,
"19.up_proj": 2762,
"19.down_proj": 2388,
"20.q_proj": 2184,
"20.k_proj": 2184,
"20.v_proj": 2184,
"20.o_proj": 1638,
"20.gate_proj": 2762,
"20.up_proj": 2762,
"20.down_proj": 2388,
"21.q_proj": 2184,
"21.k_proj": 2184,
"21.v_proj": 2184,
"21.o_proj": 1638,
"21.gate_proj": 2762,
"21.up_proj": 2762,
"21.down_proj": 2388,
"22.q_proj": 2184,
"22.k_proj": 2184,
"22.v_proj": 2184,
"22.o_proj": 1638,
"22.gate_proj": 2762,
"22.up_proj": 2762,
"22.down_proj": 2388,
"23.q_proj": 2184,
"23.k_proj": 2184,
"23.v_proj": 2184,
"23.o_proj": 1638,
"23.gate_proj": 2762,
"23.up_proj": 2762,
"23.down_proj": 2388,
"24.q_proj": 2184,
"24.k_proj": 2184,
"24.v_proj": 2184,
"24.o_proj": 1638,
"24.gate_proj": 2762,
"24.up_proj": 2762,
"24.down_proj": 2388,
"25.q_proj": 2184,
"25.k_proj": 2184,
"25.v_proj": 2184,
"25.o_proj": 1638,
"25.gate_proj": 2762,
"25.up_proj": 2762,
"25.down_proj": 2388,
"26.q_proj": 2184,
"26.k_proj": 2184,
"26.v_proj": 2184,
"26.o_proj": 1638,
"26.gate_proj": 2762,
"26.up_proj": 2762,
"26.down_proj": 2388,
"27.q_proj": 2184,
"27.k_proj": 2184,
"27.v_proj": 2184,
"27.o_proj": 1638,
"27.gate_proj": 2762,
"27.up_proj": 2762,
"27.down_proj": 2388,
"28.q_proj": 2184,
"28.k_proj": 2184,
"28.v_proj": 2184,
"28.o_proj": 1638,
"28.gate_proj": 2762,
"28.up_proj": 2762,
"28.down_proj": 2388,
"29.q_proj": 2184,
"29.k_proj": 2184,
"29.v_proj": 2184,
"29.o_proj": 1638,
"29.gate_proj": 2762,
"29.up_proj": 2762,
"29.down_proj": 2388,
"30.q_proj": 2184,
"30.k_proj": 2184,
"30.v_proj": 2184,
"30.o_proj": 1638,
"30.gate_proj": 2762,
"30.up_proj": 2762,
"30.down_proj": 2388,
"31.q_proj": 2184,
"31.k_proj": 2184,
"31.v_proj": 2184,
"31.o_proj": 1638,
"31.gate_proj": 2762,
"31.up_proj": 2762,
"31.down_proj": 2388
},
"stage": "basis_recovery_finetuned",
"recovery_finetune": {
"protocol": "basis_sharing_coefficient_lora",
"script": "gitclones/SVD-LLM/utils/LoRA.py",
"data_path": "alpaca-cleaned (/workspace/svd_safety/data/alpaca_cleaned.json)",
"epochs": 2,
"learning_rate": 0.0001,
"batch_size": 64,
"micro_batch_size": 4,
"cutoff_len": 256,
"val_set_size": 2000,
"max_train_samples": "all",
"precision": "bf16 (upstream hard-codes fp16; see run_with_torch_compat)",
"is_methods_own_stage2": false,
"note": "Basis Sharing publishes its own LoRA (lora.py: wikitext, batch 1, peft default q/v coefficient targets). This cell uses the project's SVD-LLM alpaca recipe so recovery data is constant across compressors, on all seven coefficient types and never on a basis.",
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"lora_targets": [
"down_proj",
"gate_proj",
"k_proj",
"o_proj",
"q_proj",
"up_proj",
"v_proj"
],
"trained": "Coefficient modules only; shared and private bases frozen",
"merge": "C' = C + (alpha/r) B_lora A_lora, then W = C' @ basis",
"mean_rel_delta_coefficient": 0.023527406085382348
},
"param_fraction_is_upper_bound": false,
"rel_delta_coefficient": {
"0.down_proj": 0.01414614636451006,
"0.gate_proj": 0.06271173059940338,
"0.up_proj": 0.047500111162662506,
"0.k_proj": 0.028756480664014816,
"0.o_proj": 0.012572512030601501,
"0.q_proj": 0.030865762382745743,
"0.v_proj": 0.01895243488252163,
"1.down_proj": 0.00678357295691967,
"1.gate_proj": 0.04598936438560486,
"1.up_proj": 0.040169041603803635,
"1.k_proj": 0.014593809843063354,
"1.o_proj": 0.01309526152908802,
"1.q_proj": 0.013989835046231747,
"1.v_proj": 0.014499799348413944,
"2.down_proj": 0.013051525689661503,
"2.gate_proj": 0.040834009647369385,
"2.up_proj": 0.03902892395853996,
"2.k_proj": 0.01710350811481476,
"2.o_proj": 0.015589666552841663,
"2.q_proj": 0.015468768775463104,
"2.v_proj": 0.021661514416337013,
"3.down_proj": 0.013177504763007164,
"3.gate_proj": 0.03536074981093407,
"3.up_proj": 0.03635535389184952,
"3.k_proj": 0.016993820667266846,
"3.o_proj": 0.01536262035369873,
"3.q_proj": 0.01714141294360161,
"3.v_proj": 0.02327631041407585,
"4.down_proj": 0.012231411412358284,
"4.gate_proj": 0.03274000436067581,
"4.up_proj": 0.03400220721960068,
"4.k_proj": 0.018720826134085655,
"4.o_proj": 0.015074113383889198,
"4.q_proj": 0.01642892323434353,
"4.v_proj": 0.020959772169589996,
"5.down_proj": 0.01268017292022705,
"5.gate_proj": 0.03295743092894554,
"5.up_proj": 0.032916516065597534,
"5.k_proj": 0.016637030988931656,
"5.o_proj": 0.014721281826496124,
"5.q_proj": 0.015442839823663235,
"5.v_proj": 0.020913822576403618,
"6.down_proj": 0.01286441832780838,
"6.gate_proj": 0.03249712288379669,
"6.up_proj": 0.02966822311282158,
"6.k_proj": 0.017795151099562645,
"6.o_proj": 0.014706257730722427,
"6.q_proj": 0.015420209616422653,
"6.v_proj": 0.019306669011712074,
"7.down_proj": 0.012474670074880123,
"7.gate_proj": 0.034134846180677414,
"7.up_proj": 0.03099922463297844,
"7.k_proj": 0.016013693064451218,
"7.o_proj": 0.014082911424338818,
"7.q_proj": 0.014865195378661156,
"7.v_proj": 0.020185057073831558,
"8.down_proj": 0.012424257583916187,
"8.gate_proj": 0.0344713032245636,
"8.up_proj": 0.03093084692955017,
"8.k_proj": 0.01666663959622383,
"8.o_proj": 0.01538773626089096,
"8.q_proj": 0.015956632792949677,
"8.v_proj": 0.018667476251721382,
"9.down_proj": 0.012787723913788795,
"9.gate_proj": 0.030635781586170197,
"9.up_proj": 0.03066754899919033,
"9.k_proj": 0.01625077985227108,
"9.o_proj": 0.016457993537187576,
"9.q_proj": 0.015760647132992744,
"9.v_proj": 0.01802702806890011,
"10.down_proj": 0.013270135037600994,
"10.gate_proj": 0.03306710347533226,
"10.up_proj": 0.031808458268642426,
"10.k_proj": 0.016889700666069984,
"10.o_proj": 0.014808307401835918,
"10.q_proj": 0.016796549782156944,
"10.v_proj": 0.016769904643297195,
"11.down_proj": 0.01278386078774929,
"11.gate_proj": 0.032750505954027176,
"11.up_proj": 0.03186879679560661,
"11.k_proj": 0.018679214641451836,
"11.o_proj": 0.01585085503757,
"11.q_proj": 0.017030365765094757,
"11.v_proj": 0.016336703673005104,
"12.down_proj": 0.013383959420025349,
"12.gate_proj": 0.03204452618956566,
"12.up_proj": 0.031297486275434494,
"12.k_proj": 0.018261687830090523,
"12.o_proj": 0.015123218297958374,
"12.q_proj": 0.018485387787222862,
"12.v_proj": 0.01751946099102497,
"13.down_proj": 0.01318136602640152,
"13.gate_proj": 0.033618710935115814,
"13.up_proj": 0.03187503293156624,
"13.k_proj": 0.019344637170433998,
"13.o_proj": 0.01476432103663683,
"13.q_proj": 0.01824866235256195,
"13.v_proj": 0.018521904945373535,
"14.down_proj": 0.013778213411569595,
"14.gate_proj": 0.0355733186006546,
"14.up_proj": 0.03251698985695839,
"14.k_proj": 0.019346466287970543,
"14.o_proj": 0.015444910153746605,
"14.q_proj": 0.01895812153816223,
"14.v_proj": 0.01812027022242546,
"15.down_proj": 0.013836321420967579,
"15.gate_proj": 0.033386047929525375,
"15.up_proj": 0.03200124576687813,
"15.k_proj": 0.020122358575463295,
"15.o_proj": 0.016066821292042732,
"15.q_proj": 0.019362781196832657,
"15.v_proj": 0.017482535913586617,
"16.down_proj": 0.014055527746677399,
"16.gate_proj": 0.03231862932443619,
"16.up_proj": 0.03309623897075653,
"16.k_proj": 0.022296005859971046,
"16.o_proj": 0.015972847118973732,
"16.q_proj": 0.02107972837984562,
"16.v_proj": 0.022093387320637703,
"17.down_proj": 0.01484985463321209,
"17.gate_proj": 0.03461271524429321,
"17.up_proj": 0.035233646631240845,
"17.k_proj": 0.02109636552631855,
"17.o_proj": 0.016126474365592003,
"17.q_proj": 0.0208114143460989,
"17.v_proj": 0.02159978821873665,
"18.down_proj": 0.015112471766769886,
"18.gate_proj": 0.03760548308491707,
"18.up_proj": 0.03554264083504677,
"18.k_proj": 0.022595606744289398,
"18.o_proj": 0.01604403182864189,
"18.q_proj": 0.020856209099292755,
"18.v_proj": 0.02414970099925995,
"19.down_proj": 0.01540366467088461,
"19.gate_proj": 0.03634091094136238,
"19.up_proj": 0.0359489843249321,
"19.k_proj": 0.02256142720580101,
"19.o_proj": 0.017474617809057236,
"19.q_proj": 0.02237251214683056,
"19.v_proj": 0.022501647472381592,
"20.down_proj": 0.016096290200948715,
"20.gate_proj": 0.0366799496114254,
"20.up_proj": 0.037445854395627975,
"20.k_proj": 0.02196851558983326,
"20.o_proj": 0.017981523647904396,
"20.q_proj": 0.023315971717238426,
"20.v_proj": 0.023550692945718765,
"21.down_proj": 0.015590106137096882,
"21.gate_proj": 0.038024019449949265,
"21.up_proj": 0.039658188819885254,
"21.k_proj": 0.022505933418869972,
"21.o_proj": 0.017578478902578354,
"21.q_proj": 0.021982474252581596,
"21.v_proj": 0.024720672518014908,
"22.down_proj": 0.015707720071077347,
"22.gate_proj": 0.03906581178307533,
"22.up_proj": 0.04101793095469475,
"22.k_proj": 0.022656260058283806,
"22.o_proj": 0.01720021851360798,
"22.q_proj": 0.0214886162430048,
"22.v_proj": 0.02467689849436283,
"23.down_proj": 0.015716182067990303,
"23.gate_proj": 0.041364703327417374,
"23.up_proj": 0.041586700826883316,
"23.k_proj": 0.021899448707699776,
"23.o_proj": 0.016503864899277687,
"23.q_proj": 0.022801637649536133,
"23.v_proj": 0.0233846977353096,
"24.down_proj": 0.01567365787923336,
"24.gate_proj": 0.04018592834472656,
"24.up_proj": 0.04268543794751167,
"24.k_proj": 0.022960152477025986,
"24.o_proj": 0.015794076025485992,
"24.q_proj": 0.024518316611647606,
"24.v_proj": 0.024992572143673897,
"25.down_proj": 0.01578501984477043,
"25.gate_proj": 0.0379466786980629,
"25.up_proj": 0.04037538915872574,
"25.k_proj": 0.024068159982562065,
"25.o_proj": 0.017214398831129074,
"25.q_proj": 0.02083326317369938,
"25.v_proj": 0.026051005348563194,
"26.down_proj": 0.014912595972418785,
"26.gate_proj": 0.039050161838531494,
"26.up_proj": 0.03992636501789093,
"26.k_proj": 0.02621399611234665,
"26.o_proj": 0.014508343301713467,
"26.q_proj": 0.02345091477036476,
"26.v_proj": 0.02503756619989872,
"27.down_proj": 0.015143311582505703,
"27.gate_proj": 0.038240697234869,
"27.up_proj": 0.03937762975692749,
"27.k_proj": 0.021424096077680588,
"27.o_proj": 0.01971602812409401,
"27.q_proj": 0.022661816328763962,
"27.v_proj": 0.02328498102724552,
"28.down_proj": 0.01482220645993948,
"28.gate_proj": 0.04094494506716728,
"28.up_proj": 0.03843633458018303,
"28.k_proj": 0.02151227742433548,
"28.o_proj": 0.01504539605230093,
"28.q_proj": 0.021972497925162315,
"28.v_proj": 0.023144179955124855,
"29.down_proj": 0.015450587496161461,
"29.gate_proj": 0.03961341828107834,
"29.up_proj": 0.039220333099365234,
"29.k_proj": 0.02610217034816742,
"29.o_proj": 0.014716818928718567,
"29.q_proj": 0.0211815033107996,
"29.v_proj": 0.022618116810917854,
"30.down_proj": 0.012776860035955906,
"30.gate_proj": 0.038771044462919235,
"30.up_proj": 0.03742718696594238,
"30.k_proj": 0.0264947060495615,
"30.o_proj": 0.015689056366682053,
"30.q_proj": 0.02465638890862465,
"30.v_proj": 0.018169045448303223,
"31.down_proj": 0.010587853379547596,
"31.gate_proj": 0.03626640886068344,
"31.up_proj": 0.034017015248537064,
"31.k_proj": 0.03141297027468681,
"31.o_proj": 0.01466062106192112,
"31.q_proj": 0.0292032640427351,
"31.v_proj": 0.01933012343943119
}
}