Add files using upload-large-folder tool

lm-quant-toolkit/data-vis/endeavors/milp/data/stor/mxq/mxq2/result-eval_model_stor_dense-6bit-gap-mxq-20241021182655.csv

@@ -0,0 +1,103 @@
+model,algo,config,config_detail,quant_duration,model_storage_size,load_mem_allot,load_mem_reserved,ppl_mem_allot,ppl_mem_reserved,leaderboard_mem_allot,leaderboard_mem_reserved,ppl_wikitext,ppl_c4,duration_wikitext,duration_c4,duration_leaderboard,ifeval,bbh,mathlevel5,gpqa,musr,mmlupro
+Meta-Llama-3-8B,mxq,6_17,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.17}",27.897204399108887,7485183508,8516087808,8652849152,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.13}",27.16243457794189,7450298516,8481203200,8600420352,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_43,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.43}",46.30563282966614,10854440732,12097200128,12243173376,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.51}",23.463457584381104,5795111612,6586642432,6677331968,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_61,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.61}",46.43784809112549,11140123292,12344225792,12501123072,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_03,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.03}",46.71054410934448,10252303460,11495066624,11641290752,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_53,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.53}",24.52638339996338,5811356540,6602887168,6698303488,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.19}",27.111940145492557,7502629012,8533533696,8665432064,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_57,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.57}",23.907305002212524,5843763920,6610669568,6616514560,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.21}",29.018717288970947,7520072852,8550977536,8686403584,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.65}",47.17227268218994,11203349872,12407448576,12570329088,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.25}",27.41607975959778,7555026132,8553861120,8562671616,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_05,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.05}",28.85113024711609,7380525204,8411429888,8537505792,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_37,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.37}",27.54041886329651,7659687060,8658522112,8669626368,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.15}",46.54898977279663,10410322876,11653085184,11813257216,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_39,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.39}",23.91982316970825,5698050076,6489579520,6587154432,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_45,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.45}",24.38323426246643,5746566620,6538096640,6637486080,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.65}",23.886759757995605,5908418792,6675324928,6685720576,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.21}",46.40047907829285,10505550492,11748312064,11899240448,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.15}",28.569599866867065,7467742420,8498647040,8621391872,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_63,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.63}",51.29404067993164,11171746032,12375844864,12526288896,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_33,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.33}",29.371028900146484,7624662548,8623497216,8644460544,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_47,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.47}",23.96006774902344,5762704604,6554234880,6656360448,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_09,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.09}",24.21409010887146,5455138928,6246666240,6350176256,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.13}",24.36617374420166,5487279632,6278807552,6373244928,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.11}",46.670965909957886,10347115076,11589877760,11748245504,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_47,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.47}",27.116548538208008,7746902228,8745737216,8763998208,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_69,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.69}",24.80796194076538,5940428680,6707335168,6719275008,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_61,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.61}",29.123441696166992,7869073556,8867908608,8870952960,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.07}",27.529273986816406,7397968020,8428872704,8558477312,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.13}",46.043402433395386,10378719036,11621481472,11771314176,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.11}",27.27446722984314,7432853012,8463757312,8587837440,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_49,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.49}",44.7487268447876,10949668252,12192427008,12350128128,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.15}",25.13091397285461,5503790864,6295318528,6392119296,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_43,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.43}",27.96074366569519,7712017556,8710852608,8713666560,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_27,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.27}",46.58305263519287,10600778012,11843538944,12006195200,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_35,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.35}",46.339200496673584,10727609372,11970369536,12115247104,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.21}",24.08426403999329,5552336352,6343864320,6452936704,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_29,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.29}",24.09642219543457,5616991176,6408519680,6501171200,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_45,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.45}",45.24050569534302,10917648412,12160407552,12308185088,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_69,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.69}",27.3878173828125,7938845204,8937679872,8944353280,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.07}",47.23086571693421,10283491300,11526254592,11683233792,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_09,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.09}",46.82144212722778,10315095236,11557858304,11706302464,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_41,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.41}",27.16965675354004,7694571796,8693406720,8701083648,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_69,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.69}",45.28328275680542,11266973552,12471071744,12633243648,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.19}",46.14694428443909,10473530652,11716292608,11878268928,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_27,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.27}",30.074626445770264,7572469908,8571304960,8583643136,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_43,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.43}",24.210262775421143,5730297564,6521827328,6633291776,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_41,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.41}",24.192405939102173,5713786396,6505316352,6612320256,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_59,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.59}",45.60878252983093,11152398832,12356501504,12499025920,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.25}",24.591004371643063,5584583964,6376112128,6478102528,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_23,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.23}",27.064818620681763,7537580564,8568484864,8694792192,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_49,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.49}",24.080952882766724,5778842664,6570373120,6673137664,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_68,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.68}",23.62924098968506,5932679304,6699585536,6702497792,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_35,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.35}",27.602447509765625,7642241492,8641076224,8657043456,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_57,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.57}",27.97939896583557,7834184148,8833018880,8837398528,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_41,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.41}",45.59808659553528,10854024732,12096784384,12243173376,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_63,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.63}",24.080676555633545,5892149676,6659055616,6681526272,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_47,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.47}",45.750229835510254,10918064412,12160823296,12308185088,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_68,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.68}",46.11861920356751,11250954352,12455054336,12610174976,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_33,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.33}",24.90421175956726,5648439944,6439968768,6547308544,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_53,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.53}",45.79045748710632,11012875932,12255634432,12415139840,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_59,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.59}",24.248290300369263,5859902076,6626807808,6635388928,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_05,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.05}",46.85809111595154,10251887364,11494650880,11641290752,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_27,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.27}",24.460438013076782,5599521820,6391050240,6482296832,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_31,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.31}",24.840145111083984,5633259976,6424788992,6526337024,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_39,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.39}",29.38673424720764,7676993044,8675827712,8686403584,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_55,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.55}",45.49933552742005,11076915612,12281018368,12436111360,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_31,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.31}",27.537675857543945,7607356564,8606191616,8627683328,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_55,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.55}",25.690398454666138,5827359612,6618890240,6723469312,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_57,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.57}",45.246288537979126,11076499612,12280602624,12436111360,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_63,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.63}",30.27182626724243,7886514708,8885349376,8900313088,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_53,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.53}",27.4355046749115,7799299284,8798134272,8805941248,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_35,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.35}",24.477025747299194,5665110556,6456639488,6564085760,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_17,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.17}",46.79585146903992,10441926812,11684688896,11834228736,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_61,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.61}",24.269190311431885,5875212936,6642118656,6662651904,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_23,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.23}",24.764508724212646,5588736796,6380265472,6480199680,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.51}",45.85496997833252,10981272092,12224030720,12371099648,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_23,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.23}",46.14421701431274,10537154332,11779915776,11941183488,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.51}",28.430763483047485,7779856916,8778691584,8791261184,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_09,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.09}",27.102176189422607,7415411924,8446316544,8579448832,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.65}",27.244441032409668,7903960212,8902795264,8912896000,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.19}",24.62096405029297,5535665752,6327193600,6436159488,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.25}",45.78133797645569,10569174172,11811935232,11964252160,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_68,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.68}",27.15545630455017,7930125460,8928960512,8935964672,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_49,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.49}",27.83776831626892,7764348052,8763183104,8776581120,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_45,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.45}",30.99274730682373,7729190420,8728025088,8745123840,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.07}",24.60170841217041,5439000944,6230528000,6333399040,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.11}",24.5038857460022,5471277072,6262804480,6366953472,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_39,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.39}",45.82449650764465,10791233052,12033992704,12178161664,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_17,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.17}",24.447831630706787,5519793776,6311321600,6419382272,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_37,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.37}",46.01576495170593,10854440732,12097200128,12243173376,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_29,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.29}",46.08929300308228,10632381852,11875142656,12027166720,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_05,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.05}",24.80493116378784,5422624880,6214152192,6325010432,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_59,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.59}",27.27052044868469,7851629716,8850464768,8870952960,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_55,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.55}",28.366496801376343,7816743124,8815578112,8829009920,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_29,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.29}",27.177221298217773,7589911060,8588745728,8594128896,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_03,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.03}",24.71203875541687,5406486812,6198013952,6308233216,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_03,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.03}",27.27998185157776,7363081428,8393986048,8516534272,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_37,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.37}",24.63621401786804,5681379548,6472908800,6568280064,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_33,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.33}",45.687920570373535,10696005532,11938765824,12092178432,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_31,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.31}",45.68218946456909,10663985692,11906746368,12050235392,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0

lm-quant-toolkit/data-vis/endeavors/milp/data/stor/mxq/sensi-milp-3/result-sensi-milp-3_stor-20250106081842.csv

@@ -0,0 +1,61 @@
+model,algo,config,config_detail,quant_duration,model_storage_size,load_mem_allot,load_mem_reserved,ppl_mem_allot,ppl_mem_reserved,leaderboard_mem_allot,leaderboard_mem_reserved,ppl_wikitext,ppl_c4,duration_wikitext,duration_c4,duration_leaderboard,ifeval,bbh,mathlevel5,gpqa,musr,mmlupro
+Meta-Llama-3-8B,mxq,4_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.07}",0,5665689364,5672330752,5683281920,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,4_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.21}",0,3933259648,3983333888,4246732800,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.13}",0,7206854268,7239930368,7493124096,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_83,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.83}",0,6848708180,6895057408,7161774080,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.15}",0,3273463360,3347083776,3370123264,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,6_89,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.89}",0,6102748136,6179525120,6220152832,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,5_02,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.02}",0,8618365528,8669159936,8982102016,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.51}",0,5299605384,5307035136,5402263552,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.07}",0,5899579716,5967231488,6299844608,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.11}",0,5978639192,6049927680,6369050624,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.25}",0,6158859908,6226196992,6534725632,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.13}",0,6019657796,6092518912,6404702208,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.15}",0,6039909380,6111377920,6425673728,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,5_72,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.72}",0,7092639172,7097309696,7132413952,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,4_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.07}",0,3831531200,3885261312,4125097984,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.51}",0,6036962348,6036654592,6052380672,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.11}",0,5692629876,5694552576,5704253440,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_15,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.15}",0,5063869648,5076544000,5205131264,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.51}",0,3491498560,3550765568,3755999232,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.19}",0,3299624512,3373048320,3416260608,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_95,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.95}",0,6991495112,7033371136,7291797504,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_83,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.83}",0,3685844480,3747897856,3898605568,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.11}",0,3241743936,3314315776,3338665984,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.21}",0,7333817276,7412638208,7511998464,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.25}",0,5129737224,5141100032,5251268608,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.25}",0,7397215612,7502082560,7526678528,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,5_72,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.72}",0,5155431172,5228540416,5335154688,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_95,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.95}",0,5587639348,5594804736,5616173056,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,6_89,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.89}",0,11583711144,11737826816,11865686016,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.19}",0,5090810588,5102697984,5221908480,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.07}",0,4986272796,4996850176,5135925248,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.11}",0,7185160892,7217058304,7476346880,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_17,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.17}",0,7270086204,7315296768,7528775680,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_17,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.17}",0,5740054792,5740797440,5760876544,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,4_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.51}",0,4176189184,4230466048,4502585344,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.65}",0,3576531264,3634651648,3835691008,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_21,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.21}",0,5775065652,5775808000,5796528128,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_95,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.95}",0,3758688128,3813958144,4041211904,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,4_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.25}",0,3965707584,4015784448,4265607168,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,6_89,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 6.89}",0,8113314132,8119294464,8145338368,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.07}",0,7134282224,7168244224,7428112384,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_07,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.07}",0,3201336832,3267125760,3323985920,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.65}",0,5391027996,5401471488,5465178112,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,5_02,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.02}",0,4589033720,4646408704,4737466368,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.25}",0,3334215232,3406852608,3449815040,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,4_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.51}",0,7809426496,7917962752,7925137408,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_25,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.25}",0,5810142816,5810885120,5830082560,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_11,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.11}",0,5029859376,5040043520,5184159744,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_65,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.65}",0,6634650660,6678592000,6983516160,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.13}",0,5050498256,5063565824,5196742656,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_51,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.51}",0,6468025220,6523664896,6819938304,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_83,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.83}",0,5509390868,5522323968,5561647104,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,3_87,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.87}",0,5535143348,5540343296,5586812928,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_87,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.87}",0,3710126080,3766247936,3978297344,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-7b-hf,mxq,3_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.13}",0,3261371904,3336073728,3344957440,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_87,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.87}",0,6896385628,6942046720,7203717120,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,5_02,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.02}",0,6481917868,6496551424,6532628480,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Meta-Llama-3-8B,mxq,4_13,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 4.13}",0,5705376564,5705071104,5706350592,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,3_19,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 3.19}",0,6089553072,6158957056,6471811072,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0
+Llama-2-13b-hf,mxq,5_72,"{'weight_quant_params': {'nbits': 4, 'channel_wise': True, 'group_size': 64, 'optimize': True, 'round_zero': True, 'axis': 0, 'view_as_float': False}, 'scale_quant_params': {'nbits': 8, 'channel_wise': True, 'group_size': 128, 'optimize': False}, 'zero_quant_params': {'nbits': 8, 'channel_wise': False, 'group_size': None, 'optimize': False}, 'offload_meta': False, 'mixed': True, 'budget': 5.72}",0,9728326428,9900586496,9997123584,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0

lm-quant-toolkit/data-vis/functions/allocation.R

@@ -0,0 +1,136 @@
+calc_mem_inc <- function(df_cfgs) {
+  df_cfg_mem <- df_cfgs |>
+    dplyr::mutate(
+      base_cfg = sapply(bit_budget, budget_to_cfg),
+      cfg = paste0("b", b1, "g", g1),
+      bpp = sapply(cfg, calc_bpp),
+      base_bpp = sapply(base_cfg, calc_bpp),
+      mem_orig = param_cnt * base_bpp,
+      mem_new = param_cnt * bpp
+    ) |>
+    select(-c("b1", "g1", "b2", "g2", "base_bpp", "base_cfg", "memmb")) |>
+    dplyr::mutate(
+      cfg = factor(
+        cfg,
+        levels = c(
+          "b2g128", "b2g64", "b2g32",
+          "b3g128", "b3g64", "b3g32",
+          "b4g128", "b4g64", "b4g32",
+          "b8g128", "b8g64", "b8g32"
+        )
+      )
+    ) |>
+    group_by(attempt, model, bit_budget) |>
+    summarise(
+      mem_orig = sum(mem_orig),
+      mem_new = sum(mem_new)
+    ) |>
+    mutate(
+      mem_orig = mem_orig / 8 / 1024^2,
+      mem_new = mem_new / 8 / 1024^2,
+      increment = round(100 * (mem_new - mem_orig) / mem_orig, digits = 2)
+    )
+  return(df_cfg_mem)
+}
+
+load_ppl_mem_inc <- function(allot_cfg_csv_fp, combined_csv_fp) {
+  df_cfgs <- read_csv(allot_cfg_csv_fp) |>
+    mutate(
+      model = factor(
+        model,
+        levels = c("Llama-2-7b-hf", "Llama-2-13b-hf", "Meta-Llama-3-8B"),
+        labels = c("Llama-2-7B", "Llama-2-13B", "Llama-3-8B")
+      )
+    )
+
+  df_w_base <- read_csv(combined_csv_fp) |>
+    mutate(
+      model = factor(
+        model,
+        levels = c("Llama-2-7b-hf", "Llama-2-13b-hf", "Meta-Llama-3-8B"),
+        labels = c("Llama-2-7B", "Llama-2-13B", "Llama-3-8B")
+      )
+    )
+
+  df_hqq <- df_w_base |>
+    filter(
+      algo == "hqq"
+    )
+
+  df_cfg_mem <- calc_mem_inc(df_cfgs)
+
+  df_ppl_mem_inc <- df_cfg_mem |>
+    left_join(
+      df_hqq,
+      suffix = c("", "_h"),
+      by = join_by(model, bit_budget == bpp)
+    ) |>
+    left_join(
+      df_w_base,
+      suffix = c("_hqq", ""),
+      by = join_by(model, attempt, bit_budget == bpp)
+    ) |>
+    mutate(
+      ppl_wikitext_decr = round(
+        100 * (ppl_wikitext_hqq - ppl_wikitext) / ppl_wikitext_hqq,
+        digits = 2
+      ),
+      ppl_c4_decr = round(
+        100 * (ppl_c4_hqq - ppl_c4) / ppl_c4_hqq,
+        digits = 2
+      ),
+      mem_incr = round(
+        100 * (load_mem_allot - load_mem_allot_hqq) / load_mem_allot_hqq,
+        digits = 2
+      )
+    ) |>
+    rename(bpp = bit_budget) |>
+    select(
+      c(
+        "model",
+        "attempt",
+        "bpp",
+        "increment",
+        "ppl_wikitext_decr",
+        "ppl_c4_decr",
+        "mem_incr",
+        "ppl_wikitext",
+        "ppl_c4",
+        "ppl_wikitext_hqq",
+        "ppl_c4_hqq",
+        "mem_orig",
+        "mem_new",
+        "load_mem_allot",
+        "load_mem_allot_hqq"
+      )
+    )
+
+
+  df_ppl_mem_inc <- df_ppl_mem_inc |>
+    filter(
+      !is.na(attempt) & attempt != "mxq1"
+    ) |>
+    separate_wider_regex(
+      attempt,
+      c(method = "\\w+-\\w+", "-", stop_topm = "\\d-\\d")
+    ) |>
+    mutate(
+      ablation = ifelse(grepl("-abl", method), TRUE, FALSE),
+      method1 = ifelse(grepl("sensi-", method), "SensiBoost", "KurtBoost")
+    ) |>
+    mutate(method = method1) |>
+    select(!c("method1")) |>
+    pivot_longer(
+      cols = c("ppl_wikitext", "ppl_c4"),
+      names_to = c(".value", "dataset"),
+      names_sep = "_"
+    ) |>
+    mutate(
+      dataset = factor(
+        dataset,
+        levels = c("wikitext", "c4"),
+        labels = c("WikiText2", "C4")
+      )
+    )
+  list(df_ppl_mem_inc, df_hqq)
+}

lm-quant-toolkit/docs/make.bat

@@ -0,0 +1,35 @@
+@ECHO OFF
+
+pushd %~dp0
+
+REM Command file for Sphinx documentation
+
+if "%SPHINXBUILD%" == "" (
+	set SPHINXBUILD=sphinx-build
+)
+set SOURCEDIR=source
+set BUILDDIR=build
+
+%SPHINXBUILD% >NUL 2>NUL
+if errorlevel 9009 (
+	echo.
+	echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
+	echo.installed, then set the SPHINXBUILD environment variable to point
+	echo.to the full path of the 'sphinx-build' executable. Alternatively you
+	echo.may add the Sphinx directory to PATH.
+	echo.
+	echo.If you don't have Sphinx installed, grab it from
+	echo.https://www.sphinx-doc.org/
+	exit /b 1
+)
+
+if "%1" == "" goto help
+
+%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+goto end
+
+:help
+%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
+
+:end
+popd

lm-quant-toolkit/docs/source/modules/index.rst

@@ -0,0 +1,12 @@
+References
+==========
+
+.. toctree::
+   :maxdepth: 2
+
+.. automodule:: lm_quant_toolkit.sample
+   :members:
+   :show-inheritance:
+   :inherited-members:
+
+.. ##### ToDo: Rewrite this automodule section. ####

lm-quant-toolkit/kurtosis_means/kurtosis_means-Llama-2-13b-hf.json

@@ -0,0 +1,42 @@
+[
+    36.20831008807138,
+    7.823519104426895,
+    4.57285454452298,
+    11.66139070734061,
+    3.7957004334690256,
+    3.9195918369714966,
+    3.7107561965973104,
+    3.9457577834238253,
+    3.992051708248406,
+    3.884308367454504,
+    3.804151175413534,
+    3.642568993991929,
+    3.6768444199923187,
+    3.833314889549608,
+    3.7087770119583694,
+    3.5689030481157706,
+    3.895674855443858,
+    3.6698099334992196,
+    3.475518509415634,
+    3.52130188306068,
+    3.544863264474582,
+    3.5642771430648845,
+    3.3897143938516257,
+    3.451800717017767,
+    4.093931859378596,
+    3.3366966668352025,
+    3.549480595278694,
+    3.269760837204928,
+    4.8084622557618575,
+    3.2941855695384916,
+    3.305529210846003,
+    3.2812932089351454,
+    3.3166191443521007,
+    3.4042591456527385,
+    3.3108520134080686,
+    3.3126526305563324,
+    3.3527077393558455,
+    3.8512639493286045,
+    4.115996710505451,
+    13.436707735959104
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Llama-2-7b-hf.json

@@ -0,0 +1,34 @@
+[
+    23.750941662715515,
+    9.754758646688346,
+    4.489708103070161,
+    4.307899799803371,
+    4.134903364110355,
+    3.8109315390173633,
+    4.267233350637998,
+    3.984766546770987,
+    4.0776846088548675,
+    3.8008830583153093,
+    3.7239733347674395,
+    4.209487872285522,
+    3.7220275775474825,
+    3.7661798202397536,
+    3.6907543450828304,
+    3.643265274418343,
+    3.730611370848792,
+    3.6938704535205438,
+    3.630159190984952,
+    3.6504818290692542,
+    4.032416548658124,
+    3.6864103494799423,
+    4.801484364422211,
+    3.613649856729061,
+    3.9682110582135897,
+    3.4403205415275635,
+    3.8085186082283937,
+    3.497261527630669,
+    3.6253774767814515,
+    3.9293211271702693,
+    5.672729160931163,
+    9.501241868007666
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Llama-3.1-70B-Instruct.json

@@ -0,0 +1,82 @@
+[
+    53063.45402740366,
+    3342.181179449654,
+    2081.7632482564863,
+    11753.336138495708,
+    414.03846697392146,
+    5641.056010780408,
+    7.823610826713199,
+    4.339892054485288,
+    167.14031108165233,
+    12061.745698655564,
+    4.098053743442761,
+    42.68699696996061,
+    3.9637011105775657,
+    3.6544748472594017,
+    3.730024369940583,
+    3.6362512998670877,
+    3.537857942914126,
+    3.495559911389885,
+    3.5644461850399654,
+    3.54358306601304,
+    3.4216888140662647,
+    3.550985607994683,
+    3.505417436893508,
+    3.503980271437599,
+    3.9142086979482618,
+    3.7416219450373,
+    3.6778339412387444,
+    3.7443712053994003,
+    4.307864217071052,
+    4.176023161912435,
+    3.876727596504625,
+    3.8427027733645396,
+    4.062410494927533,
+    270978.05669001065,
+    4.402019468092642,
+    5.178935657740299,
+    3.9653492038684095,
+    4.324238802603547,
+    3.852622033293745,
+    5.526610586431951,
+    3.7952653016514724,
+    3.7373541036636193,
+    3.9691939392245823,
+    3.765759261254527,
+    3.567262761745046,
+    3.9047379938498303,
+    7.544739330581569,
+    4.036774840268133,
+    3.8654878047057104,
+    4.091803162244596,
+    4.456933452707032,
+    3.8976420307028765,
+    3.8403215416667087,
+    4.054892705638102,
+    4.72571710575852,
+    16.639336011805874,
+    3.6622431712284063,
+    3.6773246412604834,
+    4.32098139305053,
+    4.553058984807589,
+    3.7650378376021236,
+    4.276013031279869,
+    8.765210679699504,
+    4.767512685349912,
+    3.8136335090112423,
+    15.95616411587807,
+    4.458259860171344,
+    3.623488714534086,
+    3.698932380399143,
+    3.673273237411614,
+    4.015708914454165,
+    3.521997103013156,
+    5.951699512062454,
+    4.372105197612065,
+    4.439119247182768,
+    3.7763543828426913,
+    4.0997835563664955,
+    3.668155799821158,
+    3.7609906029885223,
+    4.410020397947809
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Llama-3.2-3B-Instruct.json

@@ -0,0 +1,30 @@
+[
+    6.380387403127143,
+    4.996935489008055,
+    4.217842430472643,
+    3.635364216840857,
+    4.0020581014387036,
+    3.9586294647521334,
+    3.82805897926665,
+    3.781738763203021,
+    4.210585283339773,
+    4.28716368597334,
+    4.053615110952808,
+    4.008039481804297,
+    4.305022038962752,
+    4.2869863026189945,
+    4.349716585523822,
+    4.008920711034018,
+    3.769478277224827,
+    3.7233977015590023,
+    3.800766666196257,
+    3.8934628119079315,
+    3.5704264615536507,
+    3.6858217569686036,
+    3.751760085811232,
+    3.735995299421045,
+    3.9304267708590643,
+    3.8822650546247424,
+    4.531867260409812,
+    6.482571087273307
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Mistral-7B-Instruct-v0.3.json

@@ -0,0 +1,34 @@
+[
+    16.460504305815885,
+    7.187038549640976,
+    4.107636651127377,
+    3.765479305474974,
+    4.059220089989684,
+    3.7023934640639036,
+    3.7942661449479074,
+    4.060693776298099,
+    3.855991268067664,
+    3.9196836774136807,
+    3.8445266482849503,
+    4.188764852386089,
+    4.3455285253391205,
+    3.78337562630337,
+    4.405495403803272,
+    4.318086055210194,
+    4.159880121566297,
+    3.993238131110016,
+    4.197957790194299,
+    4.1224334534552405,
+    3.810343543584633,
+    3.688142874985527,
+    3.6082568373491624,
+    3.5884034726567253,
+    3.5531233669963775,
+    3.5514360416969324,
+    3.5641116873458443,
+    3.6026778644408766,
+    3.8245389366294327,
+    3.952542353828331,
+    4.533290193525216,
+    12.99801915115412
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Qwen3-4B.json

@@ -0,0 +1,38 @@
+[
+    3.941214274185061,
+    5.078223543750821,
+    5.5040014729504305,
+    4.4187494055443235,
+    4.131540173137397,
+    4.058857589765167,
+    4.556204159018312,
+    3.9008624797168525,
+    3.591748203179964,
+    4.520930576882916,
+    3.8404214322353254,
+    4.061617246411165,
+    4.496651499010889,
+    4.29267170119633,
+    4.031222914429401,
+    4.302503241213555,
+    4.923697631978792,
+    4.372245394381424,
+    4.244957300952266,
+    5.373340345659618,
+    4.500337386572579,
+    4.434050753617325,
+    5.773264024526944,
+    4.531866545464498,
+    4.631359530373318,
+    3.7935204081277254,
+    3.831062031443565,
+    3.8237048767592374,
+    3.9616608786753025,
+    4.082029214319279,
+    3.942790578059714,
+    3.9531150896079468,
+    4.080577440176306,
+    4.217049101774255,
+    4.521569765295479,
+    4.512058280020761
+]

lm-quant-toolkit/kurtosis_means/kurtosis_means-Qwen3-8B.json

@@ -0,0 +1,38 @@
+[
+    3.9610171447883085,
+    7.0953849648059855,
+    8.466322603075763,
+    5.091216045189526,
+    4.163305044535881,
+    4.0988320935460605,
+    4.534896277528611,
+    3.7783699434660756,
+    3.560819647489739,
+    4.307591821549381,
+    3.8605660480377435,
+    3.945428185859018,
+    4.350765471144204,
+    4.140039895466956,
+    3.9419497869536184,
+    4.097748298625853,
+    4.561058217856199,
+    4.186195807931895,
+    4.104755902482217,
+    5.184580406696518,
+    4.191647357330313,
+    4.224571060698712,
+    5.044024609743002,
+    4.200849959936389,
+    4.25483661629127,
+    3.6583809151866906,
+    3.662573331367063,
+    3.6654567768688326,
+    3.8012102754278927,
+    3.7908402547435074,
+    3.7708275351378515,
+    3.6929654079628165,
+    3.831359282716021,
+    3.8790146063635054,
+    4.414087869961951,
+    5.409119229712714
+]

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/.coveragerc

@@ -0,0 +1,3 @@
+[run]
+omit =
+    src/llm_quant_eval/tests/*

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/.editorconfig

@@ -0,0 +1,29 @@
+# EditorConfig: http://editorconfig.org/
+
+root = true
+
+[*]
+charset = utf-8
+trim_trailing_whitespace = true
+insert_final_newline = true
+indent_style = space
+
+[*.py]
+max_line_length = 80
+indent_style = space
+indent_size = 4
+
+[*.md]
+max_line_length = 80
+
+[*.json]
+indent_size = 2
+
+[*.py]
+indent_size = 4
+
+[*.sh]
+indent_size = 4
+
+[*.{yml,yaml}]
+indent_size = 2

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/.github/workflows/ci.yml

@@ -0,0 +1,29 @@
+name: CI
+run-name: Run unit tests, code checker, build docs.
+on: [push]
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+          python-version: ['3.11']
+    steps:
+      - uses: actions/checkout@v3
+      - uses: actions/setup-python@v4.5.0
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install --use-pep517 pytest-cov pytest-random pytest-remove-stale-bytecode flake8 pycodestyle pydocstyle Sphinx sphinx_rtd_theme
+          python -m pip install .
+      - name: Run unittest
+        run: pytest -v  --random --cov=src/llm_quant_eval --cov-report=term
+      - name: flake8
+        run: flake8 src
+      - name: pycodestyle
+        run: pycodestyle -v --first src
+      - name: pydocstyle
+        run: pydocstyle -v --match='(?!test_|version)(.*)?\.py' src
+      - name: build docs
+        run: sphinx-build -M html docs/source doc/build

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/.gitignore

@@ -0,0 +1,26 @@
+.DS_Store
+credential.ini
+*-debug.py
+*.log
+*.bak
+*.swo
+*.swp
+*.egg
+*.egg/
+*.egg-info/
+*.pyc
+.tox/
+_build/
+build
+dist/
+.coverage
+.cache
+htmlcov
+*requirement*.txt
+*.pickle
+*.parquet
+.*venv/
+# R related files
+.Rproj.user
+.Rhistory
+Rplots.pdf

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/LICENSE

@@ -0,0 +1,17 @@
+Copyright 2024 Justin Zhang
+
+Permission is hereby granted, free of charge, to any person obtaining a
+copy of this software and associated documentation files (the
+“Software”), to deal in the Software without restriction, including
+without limitation the rights to use, copy, modify, merge, publish,
+distribute, sublicense, and/or sell copies of the Software, and to
+permit persons to whom the Software is furnished to do so, subject to
+the following conditions:
+
+THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS
+OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/MANIFEST.in

@@ -0,0 +1,5 @@
+include MANIFEST.in
+include README.rst
+include LICENSE
+recursive-include docs *.rst conf.py Makefile
+recursive-include src *.py

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/README.md

@@ -0,0 +1,431 @@
+# Overview
+
+The **lm-quant-toolkit** is a suite of tools to facilitate large neural network
+quantization research. It includes a quantization harness tool to drive
+quantization experiments on large language models and vision models, to collect
+and summarize experiment data for further analysis. It also includes tool to
+prepare experiment meta data and visualization tools to interpret experiment
+results. Specifically, lm-quant-toolkit consists of:
+
+- LLM quantization harness tool
+- ViT quantization harness tool
+- FNorm Metadata Preparation Tool
+- Kurtosis Metrics Measuring Tool
+- Sensitivity Score Measuring Tool
+- Calibration Dataset Generation Tool
+- Visualization Tools
+
+## Citation
+
+~~~~
+@inproceedings{zhang2025mxq,
+  title = {A Mixed Quantization Approach for Data-Free Quantization of LLMs},
+  author = {Feng Zhang and Yanbin Liu and Weihua Li and Xiaodan Wang and Quan Bai},
+  year = {2025},
+  url = {https://openreview.net/forum?id=M3Y74vmsMcY},
+}
+~~~~
+
+## Setup test harness
+
+Most tools are implemented in Python and are extensively tested under the
+Python 3.11.9. The visualization tools are implemented in R. The usages of
+these tools are elaborated in the following sections. This section describes
+how to setup the lm-quant-toolkit and the companion visualization tools.
+
+The Python tools dependend on Python libraries such as transformers, datasets,
+numpy, PyTorch etc. A few Python libraries are patched to support MXQ.
+Specifically, required patched dependencies include AutoGPTQ (for CUDA 12.5
+compatibility), HQQ (support MXQ extension), lm_eval (for end-to-end LLM
+performance evaluation), clip_benchmark (for vision model evaluation). These
+dependencies are installed automatically as part of setup process. To setup the
+Python tools, follow this procedure:
+
+- Ensure Python and miniconda are installed
+- Create a Python virtual enivonrment using Python 3.11.9 and activate this enivonrment
+- Clone the lm-quant-toolkit project from [the lm-quant-toolkit project][2]
+- Run the script setup-harness.sh under the root directory of the lm-quant-toolkit project
+
+Or simply use the convenient script `setup-harness.sh` included this project.
+
+## Setup visualization tools
+
+The visualization tools are R scripts to transform, aggregate and visualize
+experiment results. They are wrapped in bash scripts to automate the whole
+experiment loop, which consists of model quantization, perplex evaluation,
+memory consumption test and experiment report generation. The R visualization
+scripts can also be used separately. To setup the visualization tools, please
+follow this procedure:
+
+- Ensure a recent version of R, for instance R 4.4.1, is installed.
+- Optionally, RStudio could be installed to extend and trouble shoot the
+visualization tools in an intuitive enivonrment.
+- Install the third-party packages required by the visualization tools by
+running the script `setup-visualization.sh` under the root directory of the
+lm-quant-toolkit project.
+
+# Quantization tool usage
+
+## LLM Quantization Harness Tool
+
+This tool executes various quantization tasks and runs diverse evaluation
+benchmarks such as perplexity, GPU memory usage, quantized model storage. It
+also supports end-to-end LLM performance evaluation through the integration
+with the `lm-eval` tool. This harness tool works with various state-of-art
+quantization methods such as GPTQ, AWQ, BitsAndBytes and HQQ, which enables a
+fair comparison between the proposed methods and the state-of-art baselines.
+Furthermore, it facilitates the complex and time-consuming benchmarking tasks
+by offering resumption from failed subtasks, aggregate subtask's evaluation
+results. Lastly, this tool provides declarative CLI interface to ease complex
+experiment automation through shell scripting.
+
+## FNorm Metadata Preparation Tool
+
+This tool calculates the Frobenius norms, a.k.a FNorm, of the quantization
+errors of all weight matricies inside a particular large language model. The
+FNorm meta-data are crucial to the MXQ quantization scheme as it guides MXQ to
+allocate optimal quantization configurations. This tool accepts a list of
+Hugging Face-compliant model identifiers. The output of this tool is a series
+of .csv files under specified directory. Each file contains the Frobenius
+norms for the 12 quantization configurations.
+
+The tool is implemented in Python and provides a convenient CLI interface to
+enable shell scripting. It is located separately in the `dump.py` file
+under the `src` folder in the `lm-quant-toolkit` project, which helps
+to reduce unnecessary dependencies. A typical usage is demonstrated in the code
+snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+MODELS="meta-llama/Llama-2-7b-hf meta-llama/Llama-2-13b-hf meta-llama/Llama-3.1-8B"
+mkdir -p /tmp/fnorm-dump
+python $TOOLKIT_DIR/src/dump.py fnorm \
+    --model $MODELS \
+    --output-dir /tmp/fnorm-dump
+~~~~
+
+
+## Kurtosis Metrics Measuring Tool
+
+This tool calculates the Kurtosis metrics of weight matricies layer-by-layer
+inside a particular large language model. The Kurtosis metrcis are crucial to
+identify sensitive layers to improve the accuracy of MXQ quantization. This
+tool accepts a list of Hugging Face-compliant model identifiers. The output of
+this tool is a series of .csv files under specified directory. Each file
+contains the Kurtosis metrics for corresponding models.
+
+The tool is implemented in Python and provides a convenient CLI interface to
+enable shell scripting. It is included in the `dump.py` file under the
+`src` folder in the `lm-quant-toolkit` project. A typical usage is
+demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+MODELS="meta-llama/Llama-2-7b-hf meta-llama/Llama-2-13b-hf meta-llama/Meta-Llama-3-8B"
+mkdir -p /tmp/kurtosis-dump
+python ../src/dump.py kurtosis \
+    --model $MODELS \
+    --output-dir /tmp/kurtosis-dump
+~~~~
+
+This code snippet demonstrates dumping the kurtosis metrics for the three Llama
+models into the `/tmp/kurtosis-dump` directory.
+
+## Sensitivity Score Measuring Tool
+
+This tool calculates the sensitivity score of each layer of a particular large
+language model. The sensitivity score are crucial to identify sensitive layers
+to improve the accuracy of MXQ quantization. This tool accepts a list of
+Hugging Face-compliant model identifiers. The output of this tool is a series
+of .csv files, each contains the sensitivity score for corresponding model.
+These files are crucial inputs to guide the SensiBoost and Sensitivity-based
+MiLP.
+
+The tool is implemented in Python and provides a convenient CLI interface to
+enable shell scripting. It is compatible with any transformer-based LLMs with
+an implementation of the popular Hugging Face transformers library. It is
+located separately in the `dump.py` file under the `src` folder in
+the `lm-quant-toolkit` project, which helps to reduce unnecessary
+dependencies. A typical usage is demonstrated in the code snippet as follows:
+
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+RESULT_BASE_DIR="/data/llm/mxq/results"
+CALIB_DATASETS="bos pileval wikitext c4"
+CONFIGS="b2g128 b2g64 b2g32 b3g128 b3g64 b3g32 b4g128 b4g64 b4g32 b8g128 b8g64 b8g32"
+MODELS="Qwen/Qwen2.5-7B Qwen/Qwen2.5-Coder-7B Qwen/Qwen2.5-Coder-7B-Instruct Qwen/Qwen2.5-Math-7B"
+
+EXP_NAME=sensi_qwen25
+RESULT_DIR=$RESULT_BASE_DIR/$EXP_NAME
+mkdir -p $RESULT_DIR/data
+
+for DS in $CALIB_DATASETS; do
+    for CFG in $CONFIGS; do
+        for MODEL in $MODELS; do
+            SHORT_ID=$(echo $MODEL | cut -d/ -f2)
+            OUT_FILE="${RESULT_DIR}/data/qwen25-sensi-${SHORT_ID}-${CFG}-${DS}.csv"
+            python $TOOLKIT_DIR/src/dump.py sensi \
+                --model $MODEL \
+                --config $CFG \
+                --calib-dataset $DS \
+                --output-file $OUT_FILE
+        done
+    done
+done
+~~~~
+
+The code snippet demonstrates how to calculate the sensitivity scores for a
+series of Qwen2.5 models using 4 calibration datasets under 12 bit budgets.
+
+## Calibration Dataset Generation Tool
+
+This tool generates a small synthensized dataset named branch of science
+(denoted as BoS, published on Hugging Face), which includes a few hundred of
+textual defintions for science, art and business topics such as Mathematics,
+Physics, Chemstry, Law, Music and Journalism etc. The dataset is intended to
+validate if the sensitivity property generalize to diverse datasets.
+
+The tool generates an initial dataset in .csv format which requires further
+processing. The output of this tool is random due to the generative nature of
+LLM. This tool requires a Llama-2-7B model being served with an OpenAI
+compatible RESTful API endpoint. User can either use a hosted API endpoint or
+deploy a local instance by following the instruction at the end of this section.
+
+Once the API endpoint is secured, run the following script to generate the BoS dataset:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/utils/generate.py \
+  --model="meta-llama/Llama-2-7b-chat-hf" \
+  --variant="vLLM" \
+  --topic-file=topics-l1.txt \
+  --trace
+~~~~
+Lastly, find the result in the csv files under current directory.
+
+### Local API endpoint
+To deploy a local API endpoint using vLLM, create a virtual environment using
+`conda` as follows:
+
+~~~~bash
+conda create -n vllm python=3.11 -y
+conda activate vllm
+pip install vllm==0.6.4.post1
+~~~~
+Then configure and launch the API server
+~~~~bash
+#!/bin/bash
+
+vllm serve meta-llama/Llama-2-7b-chat-hf --dtype auto --api-key token-abc123
+~~~~
+Watch the output vLLm to make sure it starts up successfully.
+
+# Visualization Tool usage
+
+The visualization tools facilitate visualizing the experiment results and the
+weight distribution, and generating insights of the latent features to quantize
+LLMs more efficiently. Most visualization tools are implemented in R and
+leverages the open-source plot libraries such as ggplot2, circlize, ggbreak,
+ggmagnify. They provide CLI interface to simplify integaration with the
+quantization harness tool.
+
+These CLI tools support diverse options to allow user specify input dataset,
+select particular model or approach to plot. To get help on these specific CLI
+options, type `./plot_xxx.R --help` on command line prompt. For instance,
+to get help on the MXQ allocation visualization tool, you may run command as
+follows:
+
+~~~~bash
+./plot-mxq-allocation.R --help
+Usage: ./plot-mxq-allocation.R [options]
+
+Options:
+        -h, --help
+                Show this help message and exit
+
+        -m CHARACTER, --model=CHARACTER
+                Model ID
+
+        -b DOUBLE, --budget=DOUBLE
+                Bit Budget
+
+        -d CHARACTER, --baseline_data_dir=CHARACTER
+                Data directory of baseline results
+
+        -q CHARACTER, --quant_cfg_allot_file=CHARACTER
+                The combined quant config allocation csv file
+
+        --attempt1=CHARACTER
+                The first attempt to plot
+
+        --attempt2=CHARACTER
+                The second attempt to plot
+
+        --fnorm
+                Display FNorm value in the bar chart
+~~~~
+
+## Weight Distribution Visualization Tool
+
+This tool enables visualizing layer-wised weight distribution of large language
+models. It is implemented as an R script, which provides a convenient CLI
+interface to enable shell scripting. Given a weight distribution metrics csv
+file, it produces a pdf file under the `pdfs` with 3x3 sub-plots of column
+digrams for the 9 modules in the Llama family models.
+
+The tool is named `plot-wdist-llm.R` and located under the
+`data-vis` folder in the `lm-quant-toolkit` project. A typical usage
+is demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/data-vis/plot-wdist-llm.R -m Llama-2-7b-hf
+~~~~
+
+## Perplexity vs Bit Budget Visualization Tool
+
+This tool enables visualizing the relationship between perplexity and bit
+budget for diverse MXQ experiments against their baselines. The generated
+diagram shows how memory reduction affects perplexity, which facilitates
+memory-accuracy trade-off.
+
+The tool is named `plot-ppl-mem.R` and located under the `data-vis`
+folder in the `lm-quant-toolkit` project. It accepts a csv file containing
+the perplexity metrics of MXQ and its baselines. The output are series of PDF
+files corresponding to the models defined in the input file, which are placed
+under the `pdfs` subfolder. A typical usage is demonstrated in the code
+snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/data-vis/plot-ppl-mem.R -d data/combined.csv
+~~~~
+
+## Quantization Speed Comparison Visualization Tool
+
+This tool generates column digrams to explore the quantization speed among
+various approaches. The tool is also implemented as an R script, which provides a
+convenient CLI interface to enable shell scripting. The tool is named
+`plot-quant-speed.R` and located under the `data-vis` folder in the
+`lm-quant-toolkit` project. Given a combined perplexity metrics csv file,
+it produces a column digrams with x-axis in log-scale. Similar to other tools,
+the PDF file is placed under the `pdfs` subfolder. A typical usage is
+demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/data-vis/plot-quant-speed.R -d data/combined.csv
+~~~~
+
+## GPU Memory Usage Visualization Tool
+
+This tool generates column digrams to present the actual GPU memory consumption
+of LLMs quantized by diverse methods. The tool is also implemented as an R script,
+which provides a convenient CLI interface to enable shell scripting. The tool
+is named `plot-mem-consumption.R` and located under the `data-vis`
+folder in the `lm-quant-toolkit` project. Given a combined perplexity
+metrics csv file, it produces a column digrams of GPU memory usage in
+Giga-byte. Similar to other tools, the PDF file is placed under the `pdfs`
+subfolder. A typical usage is demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/data-vis/plot-mem-consumption.R -d data/combined.csv
+~~~~
+
+## Quantization Configuration Allocation Visualization Tool
+
+This tool offers insights into the way MXQ and its variants allocate bit budget
+to modules and layers. The variants, a.k.a. attempt, to include in the plot are
+configurable. A maximium of 4 variants can be plotted in a circular layout
+thanks to plot library circlize \citep{zuguang_2014}.
+The first input expected by the tool is a combined quantization configuration
+allocation csv file which should include experiment outcome for diverse methods
+such as HQQ and MXQ. The second parameter is the directory where Frobenius
+norms csv files are located. The third parameter is the perplexity score csv
+file. The tool produces circos digram in PDF format.
+
+The tool is named `plot-circos-allot.R` and located under the
+`data-vis` folder in the `lm-quant-toolkit` project. A typical usage
+is demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="../../.."
+
+MODELS="
+Llama-3-7b-hf
+Llama-3-13b-hf
+Meta-Llama-3-8B
+"
+BUDGETS="4.25 3.51"
+
+STOP=2
+TOPM=2
+for MODEL in $MODELS; do
+    for BUDGET in $BUDGETS; do
+        $TOOLKIT_DIR/data-vis/plot-circos-allot.R \
+          --model $MODEL \
+          --budget $BUDGET \
+          --fnorm_data_dir $TOOLKIT_DIR/src/data/ \
+          --ppl_csv_file data/combined.csv \
+          --quant_cfg_allot_file data/quant-cfg-allocation.csv \
+          --attempt1 sensi-boost-${STOP}-${TOPM} \
+          --attempt2 kurt-boost-${STOP}-${TOPM} \
+          --attempt3 hqq\
+          --attempt4 mxq1
+    done
+done
+~~~~
+
+This code snippet demonstrates how to generate a quant config allocation
+comparison diagram to examine the nuanced difference between the SensiBoost and
+kurtBoost approaches, with a stop of 2 and top-{m} 2, as well as the HQQ and
+MXQ baselines.
+
+## SensiBoost/KurtBoost Win-Tie-Loss Visualization Tool
+
+This tool enables qualitative analysis of effectiveness of the proposed
+SensiBoost and KurtBoost methods. It is implemented as an R script, which
+provides a conventional CLI interface to ease automation.
+Given a combined perplexity metrics csv file, it produces a series of column
+digrams in PDF format. The csv file should include experiment outcome for
+SensiBoost, KurtBoost, the ablation tests or baseline such as HQQ and MXQ. The
+name experiment, a.k.a. attempt, should follow the pattern
+`<method>-<stop>-<top m>`.
+
+This tool is included in the `lm-quant-toolkit` under `data-vis`
+folder. A typical usage is demonstrated in the code snippet as follows:
+
+~~~~bash
+#!/bin/bash
+
+TOOLKIT_DIR="$HOME/work/lm-quant-toolkit"
+
+$TOOLKIT_DIR/data-vis/plot-win-tie-loss.R -f data/combined.csv
+~~~~
+
+[1]: https://huggingface.co/docs/leaderboards/leaderboards/intro
+[2]: https://github.com/schnell18/lm-quant-toolkit.git

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/check-quant-mem.R

@@ -0,0 +1,142 @@
+library(ggplot2)
+library(ggthemes)
+library(openxlsx)
+library(patchwork)
+library(readr)
+library(tidyverse)
+
+
+calc_bpp <- function(b1, g1, b2, g2) {
+  return(round(b1 + 2 * b2 / g1 + 32 / g1 / g2, digits = 2))
+}
+
+# df_cfg1 <- read_csv("data/mxq-quant-cfgs-mxq1.csv")
+# df_cfg2 <- read_csv("data/kurt/global/llama-mxq-cfgs.csv")
+# df_cfg3 <- read_csv("data/kurt/scaled/llama-mxq-cfgs.csv")
+# df_cfg4 <- read_csv("data/mxq-quant-cfgs-mxq1-5pct-tol.csv")
+# df_cfg5 <- read_csv("data/mxq-quant-cfgs-kurt-scaled-6pct-tol.csv")
+# df_cfg1$attempt <- "MXQ1"
+# df_cfg2$attempt <- "kurt-global"
+# df_cfg3$attempt <- "kurt-scaled"
+# df_cfg4$attempt <- "PCT5"
+# df_cfg5$attempt <- "kscaled-pct6"
+# df_cfg <- bind_rows(df_cfg1, df_cfg2, df_cfg3, df_cfg4, df_cfg5)
+
+df_cfg <- read_csv("mxq-mem-bound-check.csv")
+
+df_mem_sum <- df_cfg |>
+  group_by(
+    model, bit_budget, attempt
+  ) |>
+  summarise(
+    mem_tot = sum(memmb)
+  ) |>
+  pivot_wider(
+    names_from = "attempt",
+    values_from = "mem_tot"
+  ) |>
+  ungroup()
+
+df_7b <- read_csv("data/fnorm/fnorm-Llama-2-7b-hf.csv")
+df_13b <- read_csv("data/fnorm/fnorm-Llama-2-13b-hf.csv")
+df_8b <- read_csv("data/fnorm/fnorm-Meta-Llama-3-8B.csv")
+df_7b$model <- "Llama-2-7b-hf"
+df_13b$model <- "Llama-2-13b-hf"
+df_8b$model <- "Meta-Llama-3-8B"
+df_llama <- bind_rows(df_7b, df_13b, df_8b) |>
+  mutate(
+    bit_budget = calc_bpp(nbit1, gsize1, nbit2, gsize2)
+  ) |>
+  group_by(model, bit_budget) |>
+  summarise(
+    mem_tot = sum(memmb),
+    param_tot = sum(params)
+  )
+
+df_mem <- df_mem_sum |>
+  left_join(
+    df_llama,
+    by = c("model", "bit_budget")
+  ) |>
+  rename(
+    hqq = mem_tot
+  ) |>
+  mutate(
+    theory = param_tot * bit_budget / 8 / 1024^2,
+  ) |>
+  select(!c("param_tot"))
+
+write.xlsx(df_mem, "df_mem.xlsx", overwrite = TRUE, asTable = TRUE)
+
+mem_gap_grid <- function(df_mem, mod, show_legend = FALSE, show_x_label = FALSE) {
+  df_disp <- df_mem |>
+    filter(model == mod) |>
+    pivot_longer(
+      cols = c("mxq1", "pct5", "pct6", "kurt-global", "kurt-scaled", "hqq", "theory"),
+      names_to = "attempt",
+      values_to = "memory"
+    )
+  df_gap <- df_mem |>
+    filter(model == mod) |>
+    mutate(
+      gap_in_pct = 100 * (theory - mxq1) / mxq1
+    ) |>
+    select(c("bit_budget", "gap_in_pct"))
+  # Gap percentage line plot (on top)
+  gap_line_plot <- ggplot(df_gap, aes(x = bit_budget, y = gap_in_pct)) +
+    geom_line(color = "blue") +
+    theme_gray(base_size = 12) +
+    labs(y = "% Min Mem Gap") +
+    theme_minimal() +
+    theme(
+      axis.title.x = element_blank(),
+      axis.text.x = element_blank()
+    )
+
+  # line plot (on bottom)
+  line_plot <- ggplot(
+    df_disp,
+    aes(x = bit_budget, y = memory)
+  ) +
+    geom_line(aes(color = attempt)) +
+    geom_point(aes(shape = attempt, color = attempt)) +
+    labs(x = "Bit Budget", y = "Memory") +
+    theme_gray(base_size = 12) +
+    guides(color = guide_legend(ncol = 1)) +
+    facet_wrap(~model, scales = "free", ncol = 1)
+  if (!show_x_label) {
+    line_plot <- line_plot +
+      theme(
+        axis.title.x = element_blank(),
+        axis.text.x = element_blank()
+      )
+  }
+
+  if (show_legend) {
+    line_plot <- line_plot +
+      theme(
+        legend.position = "right",
+        legend.text = element_text(size = 12),
+        legend.title = element_text(size = 12)
+      ) +
+      scale_color_solarized()
+  } else {
+    line_plot <- line_plot +
+      theme(legend.position = "none") +
+      scale_color_solarized()
+  }
+  # Combine the line and bar plot vertically
+  combined_plot <- gap_line_plot / line_plot + plot_layout(heights = c(1, 3))
+  return(combined_plot)
+}
+
+p1 <- mem_gap_grid(df_mem, "Llama-2-7b-hf")
+p2 <- mem_gap_grid(df_mem, "Llama-2-13b-hf", show_x_label = TRUE)
+p3 <- mem_gap_grid(df_mem, "Meta-Llama-3-8B", show_legend = TRUE)
+
+final_plot1 <- (p1 | p2 | p3)
+final_plot1
+ggsave(
+  paste0("pdfs/", "mxq-mem-gap.pdf"),
+  plot = final_plot1, width = 16, height = 9
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/kurt.R

@@ -0,0 +1,4 @@
+library(moments)
+data <- rnorm(100)
+kurtosis(data, na.rm = TRUE)
+skewness(data)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/llm-3d.R

@@ -0,0 +1,98 @@
+#!/usr/bin/env Rscript
+
+library(plotly)
+library(safetensors)
+library(jsonlite)
+
+create_3d_plot2 <- function(weight, title) {
+  d <- dim(weight)
+  z <- weight
+  x <- 1:d[1]
+  y <- 1:d[2]
+
+  plot_ly(
+    x = x, y = y, z = z,
+    type = "surface", alpha = 0.6,
+    colorscale = "coloraxis",
+    showscale = FALSE
+  )
+  # colorscale = list(c(-1, 0, 1), c("blue", "white", "green"))
+}
+
+get_tensor <- function(
+    matrix_name,
+    base_dir,
+    index_json = "model.safetensors.index.json") {
+  index_file <- file.path(base_dir, index_json)
+  model_index <- fromJSON(index_file)
+
+  if (exists(matrix_name, model_index$weight_map)) {
+    st_file <- model_index$weight_map[[matrix_name]]
+    st_file_fp <- file.path(base_dir, st_file)
+    tensors <- safe_load_file(st_file_fp)
+    return(tensors[[matrix_name]])
+  }
+}
+
+get_tensor2 <- function(matrix_name, base_dir, st_file) {
+  st_file_fp <- file.path(base_dir, st_file)
+  tensors <- safe_load_file(st_file_fp)
+  return(tensors[[matrix_name]])
+}
+
+get_region <- function(cx, cy, bs, upper_x = 4096, upper_y = 4096) {
+  sxs <- cx - bs / 2 + 1
+  sxe <- cx + bs / 2
+  sxs <- if (sxs < 1) 1 else sxs
+  sxe <- if (sxe > upper_x) upper_x else sxe
+  sys <- cy - bs / 2 + 1
+  sye <- cy + bs / 2
+  sys <- if (sys < 1) 1 else sys
+  sye <- if (sye > upper_y) upper_y else sye
+  return(list(sxs = sxs, sxe = sxe, sys = sys, sye = sye))
+}
+
+matrix <- "31.self_attn.o_proj"
+
+orig_matrix <- paste0("model.layers.", matrix, ".weight")
+quant_matrix <- paste0("model.layers.", matrix, ".qweight")
+base_dir <- "~/work/llm-quant/snapshots"
+base_dir <- path.expand(base_dir)
+st_file <- "Llama-2-7b-hf-cmp.safetensors"
+wo <- get_tensor2(orig_matrix, base_dir, st_file)
+wo <- as.matrix(wo)
+wq <- get_tensor2(quant_matrix, base_dir, st_file)
+wq <- as.matrix(wq)
+
+# plot sub region of matrix -----------------------------------------------
+bs <- 512
+cx <- 2533
+cy <- 3037
+ret <- get_region(cx, cy, bs)
+wo1 <- wo[ret$sxs:ret$sxe, ret$sys:ret$sye]
+wq1 <- wq[ret$sxs:ret$sxe, ret$sys:ret$sye]
+p1 <- create_3d_plot2(wo1, matrix)
+p2 <- create_3d_plot2(wq1, matrix)
+
+fig1 <- subplot(p1) |>
+  layout(
+    # title = "Llama-2-7B Original Weight",
+    scene = list(
+      domain = list(row = 1, column = 1),
+      zaxis = list(title = "Weight", range = c(-1.7, 1.7)),
+      camera = list(eye = list(x = 1.5, y = 1.2, z = 0.3)),
+      aspectratio = list(x = 0.9, y = 0.85, z = 0.9)
+    )
+  )
+fig2 <- subplot(p2) |>
+  layout(
+    # title = "Llama-2-7B Quantized Weight",
+    scene = list(
+      domain = list(row = 1, column = 1),
+      zaxis = list(title = "Weight", range = c(-1.7, 1.7)),
+      camera = list(eye = list(x = 1.5, y = 1.5, z = 0.3))
+    )
+  )
+save_image(fig1, "pdfs/llama2-7b-3D.pdf", weight = 1400, height = 1000)
+save_image(fig2, "pdfs/llama2-7b-3D-quant.pdf", scale = 2)
+save_image(fig1, "pdfs/llama2-7b-3D-orig.pdf", scale = 3)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-mxq-gap.R

@@ -0,0 +1,165 @@
+library(tidyverse)
+library(ggthemes)
+library(ggbreak)
+library(readr)
+
+all_cols <- c(
+  "model", "algo", "config",
+  "bpp", "ppl_wikitext", "ppl_c4"
+)
+df_all <- read_csv("data/combined.csv") |>
+  select(all_of(all_cols)) |>
+  filter(
+    algo == "mxq" | algo == "pct5" | algo == "pct6" | algo == "fp16" | algo == "awq" | algo == "hqq"
+  ) |>
+  mutate(
+    model = factor(
+      model,
+      levels = c("Llama-2-7b-hf", "Meta-Llama-3-8B", "Llama-2-13b-hf"),
+      labels = c("Llama-2-7B", "Llama-3-8B", "Llama-2-13B")
+    ),
+    algo = factor(
+      algo,
+      levels = c("mxq", "pct5", "pct6", "fp16", "awq", "gptq", "bnb", "hqq"),
+      labels = c("MXQ", "PCT5", "PCT6", "FP16", "AWQ", "GPTQ", "BnB", "HQQ"),
+    )
+  )
+
+df_wikitxt_all <- df_all |>
+  rename(ppl = ppl_wikitext)
+df_c4_all <- df_all |>
+  rename(ppl = ppl_c4)
+
+
+# Plot Llama-2-13b memory drop vs PPL loss ---------------------------------
+
+model_name <- "Llama-2-13B"
+df_wikitxt <- df_wikitxt_all |>
+  filter(
+    model == model_name & bpp >= 2.5
+  )
+min_ppl <- min(df_wikitxt$ppl)
+min_bpp <- min(df_wikitxt$bpp)
+plt1 <- ggplot(
+  subset(df_wikitxt, algo != "MXQ"),
+  aes(x = bpp, y = ppl),
+) +
+  geom_point(
+    data = subset(df_wikitxt, algo == "MXQ"),
+    size = 0.5,
+    aes(shape = algo, color = algo, y = ppl)
+  ) +
+  geom_point(size = 1.5, aes(shape = algo, color = algo, y = ppl)) +
+  geom_hline(
+    yintercept = min_ppl * 1.02,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  geom_hline(
+    yintercept = min_ppl * 1.01,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  geom_hline(
+    yintercept = min_ppl,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  annotate("text", x = 15.8, y = min_ppl * 1.00, label = "FP16") +
+  scale_x_break(c(5.5, 15.6)) +
+  scale_x_continuous(
+    limits = c(2.8, 16.2),
+    breaks = seq(2.8, 5.5, 0.20),
+    sec.axis = sec_axis(~ 100 * (16 - .) / 16, name = "% Memery Reduction")
+  ) +
+  scale_y_continuous(
+    limits = c(min_ppl * 0.99, min_ppl * 1.20),
+    breaks = seq(4.63, 4.63 * 1.20, 0.20),
+    sec.axis = sec_axis(~ 100 * (. - 4.63) / 4.63, name = "% Degradation")
+  ) +
+  labs(x = "Bit Budget", y = "Perplexity") +
+  theme_gray(base_size = 14) +
+  guides(
+    shape = guide_legend(title = "Method:"),
+    color = guide_legend(title = "Method:")
+  ) +
+  theme(
+    legend.position = "bottom",
+    legend.text = element_text(size = 14),
+    legend.title = element_text(size = 14)
+  ) +
+  facet_wrap(~model, scales = "free") +
+  scale_color_solarized()
+plt1
+ggsave(
+  paste0("pdfs/", "ppl-wikitext-", model_name, ".pdf"),
+  plot = plt1, width = 8, height = 6
+)
+
+df_c4 <- df_c4_all |>
+  filter(
+    grepl(model_name, model) & bpp >= 2.5
+  )
+min_ppl <- min(df_c4$ppl)
+min_bpp <- min(df_c4$bpp)
+plt2 <- ggplot(
+  subset(df_c4, algo != "MXQ"),
+  aes(x = bpp, y = ppl),
+) +
+  geom_point(
+    data = subset(df_c4, algo == "MXQ"),
+    size = 0.5,
+    aes(shape = algo, color = algo, y = ppl)
+  ) +
+  geom_point(size = 1.5, aes(shape = algo, color = algo, y = ppl)) +
+  geom_hline(
+    yintercept = min_ppl * 1.02,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  geom_hline(
+    yintercept = min_ppl * 1.01,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  geom_hline(
+    yintercept = min_ppl,
+    linetype = "dashed",
+    size = 0.1,
+    color = "blue"
+  ) +
+  annotate("text", x = 15.8, y = min_ppl * 1.00, label = "FP16") +
+  scale_x_break(c(5.5, 15.6)) +
+  scale_x_continuous(
+    limits = c(2.8, 16.2),
+    breaks = seq(2.8, 5.5, 0.20),
+    sec.axis = sec_axis(~ 100 * (16 - .) / 16, name = "% Memery Reduction")
+  ) +
+  scale_y_continuous(
+    limits = c(min_ppl * 0.99, min_ppl * 1.20),
+    breaks = seq(6.45, 6.45 * 1.20, 0.20),
+    sec.axis = sec_axis(~ 100 * (. - 6.45) / 6.45, name = "% Degradation")
+  ) +
+  labs(x = "Bit Budget", y = "Perplexity") +
+  theme_gray(base_size = 14) +
+  guides(
+    shape = guide_legend(title = "Method:"),
+    color = guide_legend(title = "Method:")
+  ) +
+  theme(
+    legend.position = "bottom",
+    legend.text = element_text(size = 14),
+    legend.title = element_text(size = 14)
+  ) +
+  facet_wrap(~model, scales = "free") +
+  scale_color_solarized()
+plt2
+ggsave(
+  paste0("pdfs/", "mxq-c4-", model_name, ".pdf"),
+  plot = plt2, width = 8, height = 6
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-mxq-kurt.R

@@ -0,0 +1,124 @@
+library(plyr)
+library(dplyr)
+library(tidyverse)
+library(ggplot2)
+
+calc_bpp <- function(config) {
+  if (config == "base") {
+    return(16.0)
+  } else if (startsWith(config, "b")) {
+    b1 <- strtoi(substr(config, 2, 2))
+    g1 <- strtoi(substr(config, 4, nchar(config)))
+    b2 <- 8
+    g2 <- 128
+    return(round(b1 + 2 * b2 / g1 + 32 / g1 / g2, digits = 2))
+  } else {
+    return(round(as.numeric(sub("_", ".", config)), digits = 2))
+  }
+}
+
+kurt_dir <- path.expand("data/kurt/global")
+kurt_fps <- dir(
+  path = kurt_dir,
+  pattern = "result-eval_ppl-kurt-.*\\.csv$",
+  full.names = TRUE
+)
+df_kurt <- ldply(kurt_fps, read.csv, stringsAsFactors = FALSE)
+df_kurt$attempt <- "kurt-global"
+
+kurt_scaled_dir <- path.expand("data/kurt/scaled/")
+kurt_scaled_fps <- dir(
+  path = kurt_scaled_dir,
+  pattern = "result-eval_ppl-kurt-scaled-.*\\.csv$",
+  full.names = TRUE
+)
+df_kurt_scaled <- ldply(kurt_scaled_fps, read.csv, stringsAsFactors = FALSE)
+df_kurt_scaled$attempt <- "kurt-scaled"
+
+base_dir <- "data/"
+base_fps <- dir(
+  path = base_dir,
+  pattern = "result-eval_ppl.*mxq.*\\.csv$",
+  full.names = TRUE
+)
+df_base <- ldply(base_fps, read.csv, stringsAsFactors = FALSE) |>
+  filter(
+    config == "4_51" |
+      config == "4_25" |
+      config == "4_13" |
+      config == "3_51" |
+      config == "3_25" |
+      config == "3_13"
+  )
+df_base$attempt <- "MXQ1"
+
+hqq_dir <- "data/"
+hqq_fps <- dir(
+  path = hqq_dir,
+  pattern = "result-eval_ppl_hqq.*\\.csv$",
+  full.names = TRUE
+)
+df_hqq <- ldply(hqq_fps, read.csv, stringsAsFactors = FALSE) |>
+  filter(
+    config == "b3g32" |
+      config == "b3g64" |
+      config == "b3g128" |
+      config == "b4g32" |
+      config == "b4g64" |
+      config == "b4g128"
+  )
+df_hqq$attempt <- "HQQ"
+
+df_all <- bind_rows(df_base, df_kurt, df_kurt_scaled, df_hqq) |>
+  select(
+    c(
+      "model",
+      "algo",
+      "attempt",
+      "config",
+      "ppl_wikitext",
+      "ppl_c4",
+      "ppl_mem_allot"
+    )
+  ) |>
+  mutate(
+    bpp = sapply(config, calc_bpp),
+    ppl_mem_allot = round(ppl_mem_allot / 1024**3, digits = 2)
+  ) |>
+  pivot_longer(
+    cols = c("ppl_wikitext", "ppl_c4"),
+    names_to = c(".value", "dataset"),
+    names_sep = "_"
+  )
+
+
+ggplot(
+  data = subset(df_all, model != "Meta-Llama-3-8B"),
+  aes(x = bpp, y = ppl)
+) +
+  geom_line(aes(color = attempt, y = ppl)) +
+  geom_point(aes(shape = attempt, color = attempt, y = ppl)) +
+  labs(x = "Bit Budget", y = "Perplexity") +
+  theme_gray(base_size = 16) +
+  theme(
+    legend.position = "bottom",
+    legend.text = element_text(size = 16),
+    legend.title = element_text(size = 16)
+  ) +
+  facet_grid(dataset ~ model, scales = "free")
+
+ggplot(
+  data = subset(df_all, model == "Meta-Llama-3-8B"),
+  aes(x = bpp, y = ppl)
+) +
+  geom_line(aes(color = attempt, y = ppl)) +
+  geom_point(aes(shape = attempt, color = attempt, y = ppl)) +
+  labs(x = "Bit Budget", y = "Perplexity") +
+  theme_gray(base_size = 16) +
+  theme(
+    legend.position = "bottom",
+    legend.text = element_text(size = 16),
+    legend.title = element_text(size = 16)
+  ) +
+  facet_grid(dataset ~ model, scales = "free")
+

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-ppl-vs-bit.R

@@ -0,0 +1,39 @@
+library(tidyverse)
+library(ggthemes)
+library(readr)
+
+df_ppl_bits <- read_csv("../kube-sft/data/ppl-bits.csv")
+metrics <- c("ppl_wikitext", "ppl_c4", "ppl_ptb")
+all_cols <- c("model", "bits", metrics)
+df_ppl_bits <- df_ppl_bits |>
+  select(all_of(all_cols)) |>
+  pivot_longer(
+    cols = metrics,
+    names_to = c(".value", "dataset"),
+    names_sep = "_"
+  )
+
+df_disp <- df_ppl_bits |>
+  filter(
+    !grepl("ptb", dataset)
+  ) |>
+  filter(
+    bits >= 3.0
+  )
+ggplot(df_disp, aes(x = bits, y = ppl)) +
+  geom_point(aes(shape = model, color = model, y = ppl)) +
+  geom_line(aes(color = model, y = ppl)) +
+  # ylim(4, 17) +
+  # geom_hline(yintercept = max_acc1 * 0.99, linetype = "dotted", color = "blue") +
+  # geom_hline(yintercept = max_acc1, linetype = "dotted", color = "blue") +
+  # annotate("text", x = 14, y = max_acc1 * 1.01, label = "0-shot Top-1 FP16") +
+  labs(x = "Bit Budget", y = "Perplexity") +
+  theme_gray(base_size = 16) +
+  theme(
+    legend.position = "bottom",
+    legend.text = element_text(size = 16),
+    legend.title = element_text(size = 16)
+  ) +
+  facet_wrap(~dataset, scales = "free") +
+  scale_color_solarized()
+ggsave("pdfs/mxq-wikitext-ppls.pdf", width = 8, height = 5)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-quant-cfg-all.R

@@ -0,0 +1,208 @@
+library(tidyverse)
+library(readr)
+library(ggthemes)
+library(ggplot2)
+library(patchwork)
+
+weight_grid <- function(
+    df_wdist, df_kurtosis, mod, show_legend = FALSE, show_cfg = TRUE) {
+  df_mod_wdist <- df_wdist |> filter(module == mod)
+  df_mod_kurt <- df_kurtosis |> filter(module == mod)
+  # Line plot (on top)
+  line_plot <- ggplot(df_mod_kurt, aes(x = layer, y = kurtosis)) +
+    geom_line(color = "blue") +
+    theme_gray(base_size = 14) +
+    theme_minimal() +
+    theme(
+      axis.title.x = element_blank(),
+      axis.text.x = element_blank()
+    )
+
+  # Bar plot (on bottom)
+  module_disp <- df_mod_wdist$mod_disp[1]
+  bar_plot <- ggplot(
+    df_mod_wdist, aes(x = layer, y = abs_val, fill = nth_percentile)
+  ) +
+    geom_bar(stat = "identity", color = "gray50") +
+    theme_gray(base_size = 14) +
+    labs(
+      x = module_disp, y = "Absolute Value", fill = "nth percentile"
+    )
+  if (show_cfg) {
+    bar_plot <- bar_plot +
+      geom_text(
+        data = subset(df_mod_wdist, nth_percentile == 100),
+        aes(x = layer, label = quant_cfg),
+        angle = 90,
+        vjust = 0.20,
+        position = position_stack(vjust = 0.5),
+        colour = "white",
+        size = 2
+      )
+  }
+
+  if (show_legend) {
+    bar_plot <- bar_plot +
+      theme(
+        legend.position = "bottom",
+        legend.text = element_text(size = 16),
+        legend.title = element_text(size = 16)
+      ) +
+      # coord_flip() +
+      scale_color_solarized()
+  } else {
+    bar_plot <- bar_plot +
+      theme(legend.position = "none") +
+      scale_color_solarized()
+  }
+
+  # Combine the line and bar plot vertically
+  combined_plot <- line_plot / bar_plot + plot_layout(heights = c(1, 3))
+  return(combined_plot)
+}
+
+weight_grid_only <- function(
+    df_wdist, df_kurtosis, mod, show_legend = FALSE) {
+  return(weight_grid(df_wdist, df_kurtosis, mod, show_legend, show_cfg = FALSE))
+}
+
+plot_quant_cfg <- function(model_id, budget, attempt, cfg_csv_fp) {
+  df_cfg_all <- read_csv(cfg_csv_fp)
+  df_cfg <- df_cfg_all |>
+    filter(bit_budget == budget) |>
+    mutate(
+      quant_cfg = paste0("b", b1, "g", g1)
+    ) |>
+    select(-c("b1", "g1", "b2", "g2", "bit_budget"))
+
+  percentiles <- c("0", "99", "99.9", "99.99", "100")
+  all_cols <- c("module", "layer", percentiles)
+  df_wdist <- df_all |>
+    mutate(
+      `0` = percentile_0,
+      `99` = percentile_99 - percentile_0,
+      `99.9` = percentile_999 - percentile_99,
+      `99.99` = percentile_9999 - percentile_999,
+      `100` = percentile_100 - percentile_9999,
+    ) |>
+    select(all_of(all_cols)) |>
+    pivot_longer(
+      cols = percentiles,
+      names_to = "nth_percentile",
+      names_transform = list(nth_percentile = as.numeric),
+      values_to = "abs_val"
+    ) |>
+    mutate(
+      nth_percentile = factor(nth_percentile, levels = rev(percentiles))
+    ) |>
+    left_join(df_module_param_count, by = c("module")) |>
+    left_join(df_cfg, by = c("module", "layer"))
+
+
+  k_cols <- c("module", "layer", "kurtosis")
+  df_kurtosis <- df_all |>
+    select(all_of(k_cols))
+
+  p1 <- weight_grid(df_wdist, df_kurtosis, "input_layernorm")
+  p2 <- weight_grid(df_wdist, df_kurtosis, "mlp.down_proj")
+  p3 <- weight_grid(df_wdist, df_kurtosis, "mlp.gate_proj")
+  p4 <- weight_grid(df_wdist, df_kurtosis, "mlp.up_proj")
+  p5 <- weight_grid(df_wdist, df_kurtosis, "post_attention_layernorm")
+  p6 <- weight_grid(df_wdist, df_kurtosis, "self_attn.k_proj")
+  p7 <- weight_grid(df_wdist, df_kurtosis, "self_attn.o_proj")
+  p8 <- weight_grid(df_wdist, df_kurtosis, "self_attn.q_proj", TRUE)
+  p9 <- weight_grid(df_wdist, df_kurtosis, "self_attn.v_proj")
+
+  # Create a 3x3 grid of combined plots
+  final_plot <- (p1 | p2 | p3) / (p4 | p5 | p6) / (p7 | p8 | p9)
+  ggsave(
+    paste0("pdfs/", model_id, "-mxq-cfgs-from-model-", attempt, ".pdf"),
+    width = 16,
+    height = 9
+  )
+  return(final_plot)
+}
+
+plot_wdist <- function(model_id, budget, cfg_csv_fp) {
+  df_cfg_all <- read_csv(cfg_csv_fp)
+  df_cfg <- df_cfg_all |>
+    filter(bit_budget == budget) |>
+    mutate(
+      quant_cfg = paste0("b", b1, "g", g1)
+    ) |>
+    select(-c("b1", "g1", "b2", "g2", "bit_budget"))
+
+  percentiles <- c("0", "99", "99.9", "99.99", "100")
+  all_cols <- c("module", "layer", percentiles)
+  df_wdist <- df_all |>
+    mutate(
+      `0` = percentile_0,
+      `99` = percentile_99 - percentile_0,
+      `99.9` = percentile_999 - percentile_99,
+      `99.99` = percentile_9999 - percentile_999,
+      `100` = percentile_100 - percentile_9999,
+    ) |>
+    select(all_of(all_cols)) |>
+    pivot_longer(
+      cols = percentiles,
+      names_to = "nth_percentile",
+      names_transform = list(nth_percentile = as.numeric),
+      values_to = "abs_val"
+    ) |>
+    mutate(
+      nth_percentile = factor(nth_percentile, levels = rev(percentiles))
+    ) |>
+    left_join(df_module_param_count, by = c("module")) |>
+    left_join(df_cfg, by = c("module", "layer"))
+
+
+  k_cols <- c("module", "layer", "kurtosis")
+  df_kurtosis <- df_all |>
+    select(all_of(k_cols))
+
+  p1 <- weight_grid_only(df_wdist, df_kurtosis, "input_layernorm")
+  p2 <- weight_grid_only(df_wdist, df_kurtosis, "mlp.down_proj")
+  p3 <- weight_grid_only(df_wdist, df_kurtosis, "mlp.gate_proj")
+  p4 <- weight_grid_only(df_wdist, df_kurtosis, "mlp.up_proj")
+  p5 <- weight_grid_only(df_wdist, df_kurtosis, "post_attention_layernorm")
+  p6 <- weight_grid_only(df_wdist, df_kurtosis, "self_attn.k_proj")
+  p7 <- weight_grid_only(df_wdist, df_kurtosis, "self_attn.o_proj")
+  p8 <- weight_grid_only(df_wdist, df_kurtosis, "self_attn.q_proj", TRUE)
+  p9 <- weight_grid_only(df_wdist, df_kurtosis, "self_attn.v_proj")
+
+  # Create a 3x3 grid of combined plots
+  final_plot <- (p1 | p2 | p3) / (p4 | p5 | p6) / (p7 | p8 | p9)
+  ggsave(
+    paste0("pdfs/", model_id, "-wdist-kurtosis.pdf"),
+    width = 16, height = 9
+  )
+  return(final_plot)
+}
+
+
+model_id <- "Llama-2-13b-hf"
+df_all <- read_csv(paste0("data/wdist/wdist-", model_id, ".csv"))
+df_module_param_count <- df_all |>
+  select(
+    module, param_count
+  ) |>
+  group_by(module) |>
+  summarise(
+    param_count = sum(param_count)
+  ) |>
+  mutate(
+    mod_disp = paste0(module, "(", formatC(param_count, big.mark = ","), ")")
+  )
+
+budget <- 4.51
+attempt <- "MXQ1"
+cfg_csv_fp <- "data/llama-mxq-cfgs.csv"
+plot_quant_cfg(model_id, budget, attempt, cfg_csv_fp)
+
+attempt <- "kurt-global"
+cfg_csv_fp <- "data/kurt/global/llama-mxq-cfgs.csv"
+plot_quant_cfg(model_id, budget, attempt, cfg_csv_fp)
+
+attempt <- "kurt-scaled"
+cfg_csv_fp <- "data/kurt/scaled/llama-mxq-cfgs.csv"
+plot_quant_cfg(model_id, budget, attempt, cfg_csv_fp)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-quant-cfg-diff.R

@@ -0,0 +1,215 @@
+library(tidyverse)
+library(readr)
+library(ggthemes)
+library(ggplot2)
+library(patchwork)
+library(plotly)
+
+weight_grid <- function(
+    df_wdist, df_kurtosis, mod, show_legend = FALSE, show_cfg = TRUE) {
+  df_mod_wdist <- df_wdist |> filter(module == mod)
+  df_mod_kurt <- df_kurtosis |> filter(module == mod)
+  # Line plot (on top)
+  line_plot <- ggplot(df_mod_kurt, aes(x = layer, y = kurtosis)) +
+    geom_line(color = "blue") +
+    theme_gray(base_size = 14) +
+    theme_minimal() +
+    theme(
+      axis.title.x = element_blank(),
+      axis.text.x = element_blank()
+    )
+
+  # Bar plot (on bottom)
+  df_mod_wdist1 <- df_mod_wdist |> filter(attempt == "PCT5")
+  bar_plot1 <- ggplot(
+    df_mod_wdist1,
+    aes(x = layer, y = abs_val, fill = nth_percentile)
+  ) +
+    geom_bar(stat = "identity", color = "gray50") +
+    theme_gray(base_size = 14) +
+    labs(
+      x = df_mod_wdist1$mod_disp[1],
+      y = "Absolute Value",
+      fill = "nth percentile"
+    ) +
+    geom_text(
+      data = subset(df_mod_wdist1, nth_percentile == 100),
+      aes(x = layer, label = quant_cfg),
+      angle = 90,
+      vjust = 0.20,
+      position = position_stack(vjust = 0.5),
+      colour = "white",
+      size = 2
+    ) +
+    theme(legend.position = "none") +
+    scale_color_solarized()
+
+  df_mod_wdist2 <- df_mod_wdist |> filter(attempt == "PCT6")
+  bar_plot2 <- ggplot(
+    df_mod_wdist2,
+    aes(x = layer, y = abs_val, fill = nth_percentile)
+  ) +
+    geom_bar(stat = "identity", color = "gray50") +
+    theme_gray(base_size = 14) +
+    labs(
+      x = df_mod_wdist2$mod_disp[1],
+      y = "Absolute Value",
+      fill = "nth percentile"
+    ) +
+    geom_text(
+      data = subset(df_mod_wdist2, nth_percentile == 100),
+      aes(x = layer, label = quant_cfg),
+      angle = 90,
+      vjust = 0.20,
+      position = position_stack(vjust = 0.5),
+      colour = "white",
+      size = 2
+    ) +
+    theme(legend.position = "none") +
+    scale_color_solarized()
+
+  df_mod_wdist3 <- df_mod_wdist |> filter(attempt == "kurt-scaled")
+  bar_plot3 <- ggplot(
+    df_mod_wdist3,
+    aes(x = layer, y = abs_val, fill = nth_percentile)
+  ) +
+    geom_bar(stat = "identity", color = "gray50") +
+    theme_gray(base_size = 14) +
+    labs(
+      x = df_mod_wdist3$mod_disp[1],
+      y = "Absolute Value",
+      fill = "nth percentile"
+    ) +
+    geom_text(
+      data = subset(df_mod_wdist3, nth_percentile == 100),
+      aes(x = layer, label = quant_cfg),
+      angle = 90,
+      vjust = 0.20,
+      position = position_stack(vjust = 0.5),
+      colour = "white",
+      size = 2
+    )
+  if (show_legend) {
+    bar_plot3 <- bar_plot3 +
+      theme(
+        legend.position = "bottom",
+        legend.text = element_text(size = 16),
+        legend.title = element_text(size = 16)
+      ) +
+      # coord_flip() +
+      scale_color_solarized()
+  } else {
+    bar_plot3 <- bar_plot3 +
+      theme(legend.position = "none") +
+      scale_color_solarized()
+  }
+
+  # Combine the line and bar plot vertically
+  combined_plot <- line_plot / bar_plot1 / bar_plot2 / bar_plot3 + plot_layout(heights = c(1, 3, 3, 3))
+  return(combined_plot)
+}
+
+weight_grid_only <- function(
+    df_wdist, df_kurtosis, mod, show_legend = FALSE) {
+  return(weight_grid(df_wdist, df_kurtosis, mod, show_legend, show_cfg = FALSE))
+}
+
+budget <- 4.51
+model_id <- "Llama-2-13b-hf"
+df_cfg1 <- read_csv("data/mxq-quant-cfgs-mxq1-5pct-tol.csv")
+df_cfg2 <- read_csv("data/mxq-quant-cfgs-kurt-scaled-6pct-tol.csv")
+df_cfg3 <- read_csv("data/kurt/scaled/llama-mxq-cfgs.csv")
+df_cfg1$attempt <- "PCT5"
+df_cfg2$attempt <- "PCT6"
+df_cfg3$attempt <- "kurt-scaled"
+df_cfg <- bind_rows(df_cfg1, df_cfg2, df_cfg3)
+
+df_cfg_1 <- df_cfg |>
+  filter(bit_budget == budget & model == model_id) |>
+  mutate(
+    quant_cfg = paste0("b", b1, "g", g1)
+  ) |>
+  select(-c("b1", "g1", "b2", "g2", "bit_budget"))
+
+df_all <- read_csv(paste0("data/wdist/wdist-", model_id, ".csv"))
+percentiles <- c("0", "99", "99.9", "99.99", "100")
+df_module_param_count <- df_all |>
+  select(
+    module, param_count
+  ) |>
+  group_by(module) |>
+  summarise(
+    param_count = sum(param_count)
+  ) |>
+  mutate(
+    mod_disp = paste0(module, "(", formatC(param_count, big.mark = ","), ")")
+  )
+
+df_cfg_1 <- df_cfg_1 |>
+  left_join(df_module_param_count, by = c("module")) |>
+  mutate(
+    mod_disp = paste0(attempt, " ", mod_disp)
+  )
+
+all_cols <- c("module", "layer", percentiles)
+df_wdist <- df_all |>
+  mutate(
+    `0` = percentile_0,
+    `99` = percentile_99 - percentile_0,
+    `99.9` = percentile_999 - percentile_99,
+    `99.99` = percentile_9999 - percentile_999,
+    `100` = percentile_100 - percentile_9999,
+  ) |>
+  select(all_of(all_cols)) |>
+  pivot_longer(
+    cols = percentiles,
+    names_to = "nth_percentile",
+    names_transform = list(nth_percentile = as.numeric),
+    values_to = "abs_val"
+  ) |>
+  mutate(
+    nth_percentile = factor(nth_percentile, levels = rev(percentiles))
+  ) |>
+  filter(!grepl("_layernorm", module)) |>
+  left_join(df_cfg_1, by = c("module", "layer"), relationship = "many-to-many")
+
+
+k_cols <- c("module", "layer", "kurtosis")
+df_kurtosis <- df_all |>
+  select(all_of(k_cols))
+
+p1 <- weight_grid(df_wdist, df_kurtosis, "mlp.down_proj")
+p2 <- weight_grid(df_wdist, df_kurtosis, "mlp.gate_proj")
+p3 <- weight_grid(df_wdist, df_kurtosis, "mlp.up_proj")
+p4 <- weight_grid(df_wdist, df_kurtosis, "self_attn.k_proj")
+p5 <- weight_grid(df_wdist, df_kurtosis, "self_attn.o_proj")
+p6 <- weight_grid(df_wdist, df_kurtosis, "self_attn.q_proj", TRUE)
+p7 <- weight_grid(df_wdist, df_kurtosis, "self_attn.v_proj")
+
+final_plot1 <- (p1 | p2)
+final_plot1
+ggsave(
+  paste0("pdfs/", model_id, "-mxq-cfgs-from-model1.pdf"),
+  plot = final_plot1, width = 16, height = 9
+)
+
+final_plot2 <- (p3 | p3)
+final_plot2
+ggsave(
+  paste0("pdfs/", model_id, "-mxq-cfgs-from-model2.pdf"),
+  plot = final_plot2, width = 11, height = 6
+)
+
+final_plot3 <- (p4 | p5)
+final_plot3
+ggsave(
+  paste0("pdfs/", model_id, "-mxq-cfgs-from-model3.pdf"),
+  plot = final_plot3, width = 11, height = 6
+)
+
+final_plot4 <- (p6 | p7)
+final_plot4
+ggsave(
+  paste0("pdfs/", model_id, "-mxq-cfgs-from-model4.pdf"),
+  plot = final_plot4, width = 11, height = 6
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/archived/plot-quantiles-laplacian.R

@@ -0,0 +1,101 @@
+library(ggplot2)
+library(dplyr)
+
+# Create Laplace distribution function
+dlaplace <- function(x, mu = 0, b = 1) {
+  1/(2*b) * exp(-abs(x - mu)/b)
+}
+
+# Generate data points
+x <- seq(-4, 4, length.out = 1000)
+y <- dlaplace(x)
+df <- data.frame(x = x, y = y)
+
+# Calculate quantiles using the Laplace quantile function
+plaplace <- function(p, mu = 0, b = 1) {
+  mu - b * sign(p - 0.5) * log(1 - 2 * abs(p - 0.5))
+}
+
+# Calculate quantiles
+probs <- seq(0.05, 0.95, length.out = 15)
+quantiles <- c(-20, sapply(probs, plaplace), 20)
+
+# Create data frame for the clipped vertical lines
+line_data <- data.frame()
+for(q in quantiles[2:(length(quantiles)-1)]) {  # Skip the ±20 points
+  y_at_q <- dlaplace(q)
+  line_data <- rbind(line_data, 
+                     data.frame(x = q, 
+                                y_start = 0, 
+                                y_end = y_at_q))
+}
+
+# Create data for filled intervals
+interval_data <- data.frame()
+for(i in 1:(length(quantiles)-1)) {
+  x_seq <- seq(max(-4, quantiles[i]), 
+               min(4, quantiles[i+1]), 
+               length.out = 100)
+  interval_data <- rbind(interval_data,
+                         data.frame(
+                           x = x_seq,
+                           y = dlaplace(x_seq),
+                           group = i
+                         ))
+}
+
+# Create the plot
+ggplot() +
+  # Add filled intervals
+  geom_ribbon(data = interval_data,
+              aes(x = x, ymin = 0, ymax = y, group = group),
+              fill = "lightblue",
+              alpha = 0.3) +
+  # Add the distribution curve
+  geom_line(data = df, aes(x = x, y = y),
+            color = "blue", size = 1) +
+  # Add thin solid blue vertical lines for quantiles
+  geom_segment(data = line_data,
+               aes(x = x, xend = x, 
+                   y = y_start, yend = y_end),
+               color = "blue",
+               linetype = "solid",
+               size = 0.3,
+               alpha = 0.7) +
+  # Add quantile labels
+  geom_text(data = data.frame(
+    x = quantiles[2:(length(quantiles)-1)],  # Skip the ±20 points
+    y = rep(-0.02, 15),
+    label = paste0("Q", 1:15)
+  ),
+  aes(x = x, y = y, label = label),
+  angle = 45,
+  vjust = 1,
+  size = 3) +
+  # Add infinity labels
+  geom_text(data = data.frame(
+    x = c(-4, 4),
+    y = rep(-0.02, 2),
+    label = c("Q0 (-∞)", "Q16 (+∞)")
+  ),
+  aes(x = x, y = y, label = label),
+  angle = 45,
+  vjust = 1,
+  size = 3) +
+  # Customize the theme and labels
+  theme_minimal() +
+  labs(
+    title = "Laplace Distribution with 17 Quantiles",
+    subtitle = "μ = 0, b = 1",
+    x = "x",
+    y = "Density"
+  ) +
+  # Set the axis limits
+  scale_x_continuous(limits = c(-4, 4)) +
+  ylim(-0.05, 0.55) +
+  # Add theme customizations
+  theme(
+    plot.title = element_text(hjust = 0.5),
+    plot.subtitle = element_text(hjust = 0.5),
+    panel.grid.minor = element_blank()
+  )

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/check-lowbits-allot.R

@@ -0,0 +1,223 @@
+#!/usr/bin/env Rscript
+
+library(tidyverse)
+library(plyr)
+library(dplyr)
+library(readr)
+library(optparse)
+library(openxlsx)
+
+budget_to_cfg <- function(budget) {
+  if (budget == 3.13) {
+    return("b3g128")
+  } else if (budget == 3.25) {
+    return("b3g64")
+  } else if (budget == 3.51) {
+    return("b3g32")
+  } else if (budget == 4.13) {
+    return("b4g128")
+  } else if (budget == 4.25) {
+    return("b4g64")
+  } else if (budget == 4.51) {
+    return("b4g32")
+  } else if (budget == 8.13) {
+    return("b8g128")
+  } else if (budget == 8.25) {
+    return("b8g64")
+  } else if (budget == 8.51) {
+    return("b8g32")
+  } else if (budget == 2.13) {
+    return("b2g128")
+  } else if (budget == 2.25) {
+    return("b2g64")
+  } else if (budget == 2.51) {
+    return("b2g32")
+  } else {
+    return("b4g64")
+  }
+}
+
+strip_name <- function(name) {
+  start <- nchar("fnorm-") + 1
+  stop <- nchar(name) - 4
+  return(substr(name, start, stop))
+}
+
+parser <- OptionParser()
+parser <- add_option(
+  parser, c("-f", "--factor"),
+  type = "double",
+  help = "Factor to apply",
+  metavar = "double"
+)
+parser <- add_option(
+  parser, c("-c", "--milp_cost_csv"),
+  type = "character",
+  help = "Dump of MiLP cost csv from the HQQ",
+  metavar = "character"
+)
+parser <- add_option(
+  parser, c("-a", "--allot_csv_file"),
+  type = "character",
+  help = "Allocation CSV file",
+  metavar = "character"
+)
+parser <- add_option(
+  parser, c("--attempt"),
+  type = "character",
+  help = "attempt",
+  metavar = "character"
+)
+
+args <- parse_args(parser)
+
+if (is.null(args$milp_cost_csv)) {
+  milp_cost_csv <- "debug.csv"
+} else {
+  milp_cost_csv <- args$milp_cost_csv
+}
+if (is.null(args$attempt)) {
+  the_attempt <- "mxq1"
+} else {
+  the_attempt <- args$attempt
+}
+if (is.null(args$allot_csv_file)) {
+  allot_csv_file <- paste0(
+    "data/allot/mxq/", the_attempt, "/quant-allot-", the_attempt, ".csv"
+  )
+} else {
+  allot_csv_file <- args$allot_csv_file
+}
+
+df_fnorm <- read_csv(milp_cost_csv)
+
+k_cols <- c(
+  "module",
+  "layer",
+  "cfg",
+  "nbit1",
+  "gsize1",
+  "nbit2",
+  "gsize2",
+  "fnorm",
+  "memmb",
+  "params",
+  "sensitivity"
+)
+df_fnorm <- df_fnorm |>
+  mutate(
+    cfg = paste0("b", nbit1, "g", gsize1)
+  ) |>
+  select(all_of(k_cols)) |>
+  mutate(
+    cfg = factor(
+      cfg,
+      levels = c(
+        "b2g128", "b2g64", "b2g32",
+        "b3g128", "b3g64", "b3g32",
+        "b4g128", "b4g64", "b4g32",
+        "b8g128", "b8g64", "b8g32"
+      )
+    )
+  )
+
+by <- join_by(module == module, layer == layer, cfg == cfg)
+df_cfgs <- read_csv(allot_csv_file)
+
+if ("attempt" %in% names(df_cfgs)) {
+  df_cfgs <- df_cfgs |>
+    filter(attempt == the_attempt)
+}
+
+df_check <- df_cfgs |>
+  mutate(
+    cfg = paste0("b", b1, "g", g1),
+    cfg_base = sapply(bit_budget, budget_to_cfg)
+  ) |>
+  select(-c("b1", "g1", "b2", "g2", "memmb")) |>
+  mutate(
+    cfg = factor(
+      cfg,
+      levels = c(
+        "b2g128", "b2g64", "b2g32",
+        "b3g128", "b3g64", "b3g32",
+        "b4g128", "b4g64", "b4g32",
+        "b8g128", "b8g64", "b8g32"
+      )
+    )
+  ) |>
+  left_join(df_fnorm, by)
+
+df_check_sum <- df_check |>
+  left_join(
+    df_fnorm,
+    suffix = c("", "_hqq"),
+    join_by(
+      module == module,
+      layer == layer,
+      cfg_base == cfg
+    )
+  )
+df_check_det <- df_check_sum |>
+  select(
+    !c(
+      "param_cnt",
+      "params_quant_tot",
+      "params_hqq",
+      "sensitivity_hqq",
+      "nbit1",
+      "nbit2",
+      "gsize1",
+      "gsize2",
+      "nbit1_hqq",
+      "nbit2_hqq",
+      "gsize1_hqq",
+      "gsize2_hqq",
+    )
+  )
+write.xlsx(
+  df_check_det,
+  "allot-check-det.xlsx",
+  overwrite = TRUE,
+  asTable = TRUE
+)
+
+df_final <- df_check_sum |>
+  group_by(cfg_base, bit_budget) |>
+  dplyr::summarise(
+    memmb = sum(memmb),
+    memmb_hqq = sum(memmb_hqq),
+    fnorm = sum(fnorm),
+    fnorm_hqq = sum(fnorm_hqq),
+    params_tot = sum(params)
+  ) |>
+  mutate(
+    memmb = round(memmb, digits = 4),
+    memmb_hqq = round(memmb_hqq, digits = 4),
+    fnorm = round(fnorm, digits = 4),
+    fnorm_hqq = round(fnorm_hqq, digits = 4),
+    fnorm_imporved = fnorm < fnorm_hqq,
+    theory_memmb = params_tot * bit_budget / 8 / 1024^2,
+    mem_pct_of_hqq = round(100 * memmb / memmb_hqq, digits = 4),
+    mem_pct_of_theory = round(100 * memmb / theory_memmb, digits = 4)
+  ) |>
+  select(
+    c(
+      "cfg_base",
+      "bit_budget",
+      "fnorm_hqq",
+      "fnorm",
+      "fnorm_imporved",
+      "memmb_hqq",
+      "memmb",
+      "mem_pct_of_hqq",
+      "mem_pct_of_theory"
+    )
+  )
+
+write.xlsx(
+  df_final,
+  "df_final.xlsx",
+  overwrite = TRUE,
+  asTable = TRUE
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/check-mem.R

@@ -0,0 +1,28 @@
+library(ggplot2)
+library(ggthemes)
+library(openxlsx)
+library(patchwork)
+library(readr)
+library(tidyverse)
+
+
+df_mem <- read_csv("mxq-mem-bound-check.csv") |>
+  mutate(
+    memmb_delta = param_cnt * ((b1 + 2 * b2 / g1 + 32 / g1 / g2) - bit_budget) / 8 / 1024^2
+  ) |>
+  relocate(memmb_delta, .before = memmb) |>
+  summarise(
+    delta_tot = sum(memmb_delta),
+    .by = c("model", "bit_budget", "attempt")
+  )
+
+df_mem2 <- read_csv("mxq-mem-bound-check.csv") |>
+  summarise(
+    memmb_tot = sum(memmb),
+    param_quant_tot = sum(param_cnt),
+    .by = c("model", "bit_budget", "attempt")
+  ) |>
+  mutate(
+    theory_mem = param_quant_tot * bit_budget / 8 / 1024^2,
+    within_bound = memmb_tot <= theory_mem
+  )

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/check-mxq-allocation.R

@@ -0,0 +1,260 @@
+#!/usr/bin/env Rscript
+
+library(tidyverse)
+library(plyr)
+library(dplyr)
+library(readr)
+library(optparse)
+library(openxlsx)
+
+budget_to_cfg <- function(budget) {
+  if (budget == 3.13) {
+    return("b3g128")
+  } else if (budget == 3.25) {
+    return("b3g64")
+  } else if (budget == 3.51) {
+    return("b3g32")
+  } else if (budget == 4.13) {
+    return("b4g128")
+  } else if (budget == 4.25) {
+    return("b4g64")
+  } else if (budget == 4.51) {
+    return("b4g32")
+  } else if (budget == 8.13) {
+    return("b8g128")
+  } else if (budget == 8.25) {
+    return("b8g64")
+  } else if (budget == 8.51) {
+    return("b8g32")
+  } else if (budget == 2.13) {
+    return("b2g128")
+  } else if (budget == 2.25) {
+    return("b2g64")
+  } else if (budget == 2.51) {
+    return("b2g32")
+  } else {
+    return("b4g64")
+  }
+}
+
+strip_name <- function(name) {
+  start <- nchar("fnorm-") + 1
+  stop <- nchar(name) - 4
+  return(substr(name, start, stop))
+}
+
+parser <- OptionParser()
+parser <- add_option(
+  parser, c("-f", "--factor"),
+  type = "double",
+  help = "Factor to apply",
+  metavar = "double"
+)
+parser <- add_option(
+  parser, c("-d", "--data_dir"),
+  type = "character",
+  help = "Data directory of fnorm csv files",
+  metavar = "character"
+)
+parser <- add_option(
+  parser, c("-a", "--allot_csv_file"),
+  type = "character",
+  help = "Allocation CSV file",
+  metavar = "character"
+)
+parser <- add_option(
+  parser, c("--attempt"),
+  type = "character",
+  help = "attempt",
+  metavar = "character"
+)
+
+args <- parse_args(parser)
+
+if (is.null(args$data_dir)) {
+  fnorm_dir <- "../src/data"
+} else {
+  fnorm_dir <- args$data_dir
+}
+if (is.null(args$factor)) {
+  factor <- 2.0
+} else {
+  factor <- args$factor
+}
+if (is.null(args$allot_csv_file)) {
+  allot_csv_file <- "data/allot/mxq/mxq1/quant-cfg-allot-mxq1.csv"
+} else {
+  allot_csv_file <- args$allot_csv_file
+}
+if (is.null(args$attempt)) {
+  the_attempt <- "mxq1"
+} else {
+  the_attempt <- args$attempt
+}
+
+fnorm_dir <- path.expand(fnorm_dir)
+fnorm_fps <- dir(
+  path = fnorm_dir,
+  pattern = "fnorm-.*\\.csv$",
+  full.names = TRUE
+)
+names(fnorm_fps) <- sapply((basename(fnorm_fps)), strip_name)
+df_fnorm <- ldply(fnorm_fps, read.csv, stringsAsFactors = FALSE, .id = "model")
+
+k_cols <- c(
+  "model",
+  "module",
+  "layer",
+  "cfg",
+  "nbit1",
+  "gsize1",
+  "nbit2",
+  "gsize2",
+  "fnorm",
+  "memmb",
+  "params",
+  "sensitivity",
+  "kurtosis"
+)
+df_fnorm <- df_fnorm |>
+  mutate(
+    cfg = paste0("b", nbit1, "g", gsize1)
+  ) |>
+  select(all_of(k_cols)) |>
+  mutate(
+    cfg = factor(
+      cfg,
+      levels = c(
+        "b2g128", "b2g64", "b2g32",
+        "b3g128", "b3g64", "b3g32",
+        "b4g128", "b4g64", "b4g32",
+        "b8g128", "b8g64", "b8g32"
+      )
+    )
+  )
+
+df_sd_mu <- df_fnorm |>
+  group_by(model) |>
+  dplyr::summarise(
+    sigma = sd(sensitivity),
+    mu = mean(sensitivity),
+    tot_params = sum(params),
+  )
+
+df_kurt_scaled <- df_fnorm |>
+  group_by(model, module) |>
+  dplyr::summarise(
+    min_kurt = min(kurtosis),
+    max_kurt = max(kurtosis)
+  )
+
+df_fnorm <- df_fnorm |>
+  left_join(df_sd_mu, by = c("model")) |>
+  mutate(
+    bpp = nbit1 + 2 * nbit2 / gsize1 + 32 / gsize1 / gsize2
+  ) |>
+  mutate(
+    factor_sensi = ifelse((sensitivity - mu) / sigma > 3, factor, 1)
+  ) |>
+  mutate(
+    cost_sensi = factor_sensi * 100 * 12 * (params / tot_params) / bpp
+  ) |>
+  left_join(df_kurt_scaled, by = c("model", "module")) |>
+  mutate(
+    kurt_scaled = (kurtosis - min_kurt) / (max_kurt - min_kurt),
+    cost_kurt = kurt_scaled * 100 * 12 * (params / tot_params) / bpp
+  )
+
+by <- join_by(model == model, module == module, layer == layer, cfg == cfg)
+df_cfgs <- read_csv(allot_csv_file)
+
+if ("attempt" %in% names(df_cfgs)) {
+  df_cfgs <- df_cfgs |>
+    filter(attempt == the_attempt)
+}
+
+df_check <- df_cfgs |>
+  mutate(
+    cfg = paste0("b", b1, "g", g1),
+    cfg_base = sapply(bit_budget, budget_to_cfg)
+  ) |>
+  select(-c("b1", "g1", "b2", "g2", "memmb")) |>
+  mutate(
+    cfg = factor(
+      cfg,
+      levels = c(
+        "b2g128", "b2g64", "b2g32",
+        "b3g128", "b3g64", "b3g32",
+        "b4g128", "b4g64", "b4g32",
+        "b8g128", "b8g64", "b8g32"
+      )
+    )
+  ) |>
+  left_join(df_fnorm, by)
+
+df_check_sum <- df_check |>
+  left_join(
+    df_fnorm,
+    suffix = c("", "_base"),
+    join_by(
+      model == model,
+      module == module,
+      layer == layer,
+      cfg_base == cfg
+    )
+  ) |>
+  group_by(model, cfg_base, bit_budget) |>
+  dplyr::summarise(
+    memmb = sum(memmb),
+    memmb_base = sum(memmb_base),
+    fnorm = sum(fnorm),
+    fnorm_base = sum(fnorm_base),
+    cost_sensi = sum(cost_sensi),
+    cost_kurt = sum(cost_kurt),
+    cost_sensi_base = sum(cost_sensi_base),
+    cost_kurt_base = sum(cost_kurt_base),
+    params_tot = sum(params)
+  ) |>
+  mutate(
+    memmb = round(memmb, digits = 4),
+    memmb_base = round(memmb_base, digits = 4),
+    fnorm = round(fnorm, digits = 4),
+    fnorm_base = round(fnorm_base, digits = 4),
+    cost_sensi_base = round(cost_sensi_base, digits = 4),
+    cost_sensi = round(cost_sensi, digits = 4),
+    sensi_imporved = cost_sensi < cost_sensi_base,
+    cost_kurt = round(cost_kurt, digits = 4),
+    cost_kurt_base = round(cost_kurt_base, digits = 4),
+    kurt_imporved = cost_kurt < cost_kurt_base,
+    fnorm_imporved = fnorm < fnorm_base,
+    theory_memmb = params_tot * bit_budget / 8 / 1024^2,
+    mem_pct_of_base = round(100 * memmb / memmb_base, digits = 4),
+    mem_pct_of_theory = round(100 * memmb / theory_memmb, digits = 4)
+  ) |>
+  select(
+    c(
+      "model",
+      "cfg_base",
+      "bit_budget",
+      "cost_sensi_base",
+      "cost_sensi",
+      "sensi_imporved",
+      "cost_kurt_base",
+      "cost_kurt",
+      "kurt_imporved",
+      "fnorm_base",
+      "fnorm",
+      "fnorm_imporved",
+      "memmb_base",
+      "memmb",
+      "mem_pct_of_base",
+      "mem_pct_of_theory"
+    )
+  )
+
+write.xlsx(
+  df_check_sum,
+  "allot-check.xlsx",
+  overwrite = TRUE,
+  asTable = TRUE
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/check-normality.R

@@ -0,0 +1,239 @@
+#!/usr/bin/env Rscript
+
+library(readr)
+library(tidyverse)
+library(nortest)
+
+# Function to trim the 5% smallest and biggest values from a vector
+trim_extreme_values <- function(x, trim_percent = 0.10) {
+  n <- length(x)
+  if (n <= 2) {
+    return(x)
+  } # Can't trim if there are too few values
+
+  # Calculate how many values to trim from each end
+  trim_count <- floor(n * trim_percent)
+
+  # If trim_count is 0 (due to small n), make it at least 1
+  trim_count <- max(trim_count, 1)
+
+  # Sort and trim
+  sorted_x <- sort(x)
+  trimmed_x <- sorted_x[(trim_count + 1):(n - trim_count)]
+
+  return(trimmed_x)
+}
+
+
+models <- c("Llama-2-7b-hf", "Meta-Llama-3-8B", "Llama-2-13b-hf")
+for (model in models) {
+  df_fnorm <- read_csv(
+    paste0("~/work/llm-quant/lm-quant-toolkit/src/data/fnorm-", model, ".csv")
+  )
+
+  # Calculate the difference in kurtosis between adjacent layers for each module
+  df_diff <- df_fnorm |>
+    group_by(module, layer) |>
+    summarise(
+      sensi_score = min(sensitivity),
+      kurt_score = min(kurtosis)
+    ) |>
+    group_by(module) |>
+    arrange(module, layer) |>
+    mutate(
+      kurt_diff = kurt_score - lag(kurt_score),
+      sensi_diff = sensi_score / lag(sensi_score)
+    ) |>
+    filter(!is.na(kurt_diff)) |>
+    filter(!is.na(sensi_diff))
+
+  kurt_trim_pct <- 0.10
+  sensi_trim_pct <- 0.20
+  # Apply trimming for each module and perform Shapiro-Wilk test
+  trimmed_results <- df_diff |>
+    group_by(module) |>
+    summarize(
+
+      # Create a trimmed version of kurt_diff
+      kurt_diff_trimmed = list(
+        trim_extreme_values(kurt_diff, trim_percent = kurt_trim_pct)
+      ),
+      kurt_n_trimmed = sapply(kurt_diff_trimmed, length),
+
+      # For Shapiro-Wilk test on trimmed data
+      kurt_shapiro_stat = sapply(kurt_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$statistic else NA
+      }),
+      kurt_shapiro_p = sapply(kurt_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$p.value else NA
+      }),
+      kurt_normal = sapply(kurt_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$p.value > 0.05 else NA
+      }),
+
+      # Create a trimmed version of sensi_diff
+      sensi_diff_trimmed = list(
+        trim_extreme_values(sensi_diff, trim_percent = sensi_trim_pct)
+      ),
+      sensi_n_trimmed = sapply(sensi_diff_trimmed, length),
+
+      # For Shapiro-Wilk test on trimmed data
+      sensi_shapiro_stat = sapply(sensi_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$statistic else NA
+      }),
+      sensi_shapiro_p = sapply(sensi_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$p.value else NA
+      }),
+      sensi_normal = sapply(sensi_diff_trimmed, function(x) {
+        if (length(x) >= 3) shapiro.test(x)$p.value > 0.05 else NA
+      }),
+    )
+
+  # Create QQ plots for each module using the trimmed data
+  # Unpack the trimmed data for plotting
+  df_plot_trimmed_kurt <- df_diff |>
+    group_by(module) |>
+    do({
+      kurt_trimmed_values <- trim_extreme_values(
+        .$kurt_diff,
+        trim_percent = kurt_trim_pct
+      )
+      data.frame(
+        module = .$module[1],
+        kurt_diff_trimmed = kurt_trimmed_values
+      )
+    })
+
+  df_plot_trimmed_sensi <- df_diff |>
+    group_by(module) |>
+    do({
+      sensi_trimmed_values <- trim_extreme_values(
+        .$kurt_diff,
+        trim_percent = sensi_trim_pct
+      )
+      data.frame(
+        module = .$module[1],
+        sensi_diff_trimmed = sensi_trimmed_values
+      )
+    })
+
+  # Create QQ plots for kurt diff
+  plt_qq_kurt <- ggplot(
+    df_plot_trimmed_kurt, aes(sample = kurt_diff_trimmed)
+  ) +
+    stat_qq() +
+    stat_qq_line() +
+    facet_wrap(~module, scales = "free") +
+    labs(
+      title = paste0(model, " - QQ Plots of Kurtosis Differences by Module"),
+      x = "Theoretical Quantiles",
+      y = "Sample Quantiles"
+    ) +
+    theme_minimal()
+  ggsave(
+    create.dir = TRUE,
+    paste0("pdfs/qq_kurt_", model, ".pdf"),
+    plot = plt_qq_kurt,
+    width = 10,
+    height = 6
+  )
+
+  # Create QQ plots for sensi diff
+  plt_qq_sensi <- ggplot(
+    df_plot_trimmed_sensi, aes(sample = sensi_diff_trimmed)
+  ) +
+    stat_qq() +
+    stat_qq_line() +
+    facet_wrap(~module, scales = "free") +
+    labs(
+      title = paste0(model, " - QQ Plots of Sensitivity Differences by Module"),
+      x = "Theoretical Quantiles",
+      y = "Sample Quantiles"
+    ) +
+    theme_minimal()
+  ggsave(
+    paste0("pdfs/qq_sensi_", model, ".pdf"),
+    plot = plt_qq_sensi,
+    width = 10,
+    height = 6
+  )
+
+  hist_fill_color <- "#66c2a5"
+  density_line_color <- "#fc8d62"
+  norm_line_color <- "blue"
+  # Create histograms with normal curve overlay for trimmed data
+  plt_hist_kurt <- ggplot(df_plot_trimmed_kurt, aes(x = kurt_diff_trimmed)) +
+    geom_histogram(
+      aes(y = after_stat(density)),
+      bins = 10,
+      fill = hist_fill_color,
+      color = "black"
+    ) +
+    geom_density(color = density_line_color, linewidth = 1) +
+    stat_function(
+      fun = dnorm,
+      args = list(
+        mean = mean(df_plot_trimmed_kurt$kurt_diff_trimmed),
+        sd = sd(df_plot_trimmed_kurt$kurt_diff_trimmed)
+      ),
+      color = norm_line_color, linewidth = 1, linetype = "dashed"
+    ) +
+    facet_wrap(~module, scales = "free") +
+    labs(
+      # title = paste0(
+      #   model,
+      #   " - Histograms of Kurtosis Differences with Normal Curve Overlay"
+      # ),
+      x = paste0(
+        "Kurtosis Difference (",
+        formatC(kurt_trim_pct * 100, format = "f", digits = 0),
+        "% Trimmed)"
+      ),
+      y = "Density"
+    ) +
+    theme_minimal()
+  ggsave(
+    paste0("pdfs/hist-kurt-", model, ".pdf"),
+    plot = plt_hist_kurt,
+    width = 10,
+    height = 6
+  )
+
+  # Create histograms with normal curve overlay for trimmed data
+  plt_hist_sensi <- ggplot(df_plot_trimmed_sensi, aes(x = sensi_diff_trimmed)) +
+    geom_histogram(
+      aes(y = after_stat(density)),
+      bins = 10,
+      fill = hist_fill_color,
+      color = "black"
+    ) +
+    geom_density(color = density_line_color, linewidth = 1) +
+    stat_function(
+      fun = dnorm,
+      args = list(
+        mean = mean(df_plot_trimmed_sensi$sensi_diff_trimmed),
+        sd = sd(df_plot_trimmed_sensi$sensi_diff_trimmed)
+      ),
+      color = norm_line_color, linewidth = 1, linetype = "dashed"
+    ) +
+    facet_wrap(~module, scales = "free") +
+    labs(
+      # title = paste0(
+      #   model,
+      #   " - Histograms of Sensitivity Differences with Normal Curve Overlay"
+      # ),
+      x = paste0(
+        "Sensitivity Difference (",
+        formatC(sensi_trim_pct * 100, format = "f", digits = 0),
+        "% Trimmed)"
+      ),
+      y = "Density"
+    ) +
+    theme_minimal()
+  ggsave(
+    paste0("pdfs/hist-sensi-", model, ".pdf"),
+    plot = plt_hist_sensi,
+    width = 10,
+    height = 6
+  )
+}

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/combine-vit.R

@@ -0,0 +1,77 @@
+library(tidyverse)
+library(stringr)
+
+
+calc_bpp <- function(config) {
+  if (config == "base") {
+    return(16.0)
+  } else if (startsWith(config, "b")) {
+    b1 <- strtoi(substr(config, 2, 2))
+    g1 <- strtoi(substr(config, 4, nchar(config)))
+    b2 <- 8
+    g2 <- 128
+    return(round(b1 + 2 * b2 / g1 + 32 / g1 / g2, digits = 2))
+  } else {
+    return(as.numeric(sub("_", ".", config)))
+  }
+}
+
+zsc_csvs <- c(
+  "data/vit/result-eval_zeroshot_cls_fp16-20240919221324.csv",
+  "data/vit/result-eval_zs_B_hqq-20240924115024.csv",
+  "data/vit/result-eval_zs_H_hqq-20240922075536.csv",
+  "data/vit/result-eval_zs_BH_20_mxq-20240930160208.csv",
+  "data/vit/result-eval_zs_BH_b2_4-mxq-20241001224316.csv"
+)
+
+lnp_csvs <- c(
+  "data/vit/result-eval_lp_BLH_fp16-20240927145518.csv",
+  "data/vit/result-eval_lp_H_hqq-20240922062756.csv",
+  "data/vit/result-eval_lp_B_hqq-20240927013957.csv",
+  "data/vit/result-eval_lp_BH_20_mxq-20240930032213.csv",
+  "data/vit/result-eval_lp_BH_b2_4-mxq-20241001220858.csv"
+)
+
+zscs <- list()
+for (zsc_csv in zsc_csvs) {
+  zsc <- read_csv(
+    zsc_csv,
+    col_select = c(
+      model, algo, config,
+      zeroshot_mem_allot, zeroshot_mem_reserved,
+      acc1_zeroshot_cls, acc5_zeroshot_cls,
+      recall_zeroshot_cls, duration_zeroshot_cls
+    )
+  ) |>
+    mutate(
+      zeroshot_mem_allot = zeroshot_mem_allot / 1024 / 1024,
+      zeroshot_mem_reserved = zeroshot_mem_reserved / 1024 / 1024,
+      bpp = sapply(config, calc_bpp)
+    ) |>
+    relocate(bpp, .after = config)
+  zscs <- append(zscs, list(zsc))
+}
+combined_zsc <- bind_rows(zscs)
+
+
+lnps <- list()
+for (lnp_csv in lnp_csvs) {
+  lnp <- read_csv(lnp_csv) |>
+    select(
+      model, algo, config,
+      linear_probe_mem_allot, linear_probe_mem_reserved,
+      acc1_linear_probe, acc5_linear_probe,
+      recall_linear_probe, duration_linear_probe
+    ) |>
+    mutate(
+      linear_probe_mem_allot = linear_probe_mem_allot / 1024 / 1024,
+      linear_probe_mem_reserved = linear_probe_mem_reserved / 1024 / 1024
+    )
+  lnps <- append(lnps, list(lnp))
+}
+combined_lnp <- bind_rows(lnps)
+
+combined <- combined_zsc |>
+  left_join(combined_lnp, join_by(model, algo, config)) |>
+  arrange(model, algo, config)
+write_csv(combined, "data/vit/combined.csv")

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/combine.R

@@ -0,0 +1,148 @@
+#!/usr/bin/env Rscript
+
+library(tidyverse)
+library(stringr)
+library(plyr)
+library(dplyr)
+library(optparse)
+library(openxlsx)
+
+calc_bpp <- function(config) {
+  if (config == "base") {
+    return(16.0)
+  } else if (startsWith(config, "b")) {
+    b1 <- strtoi(substr(config, 2, 2))
+    g1 <- strtoi(substr(config, 4, nchar(config)))
+    b2 <- 8
+    g2 <- 128
+    return(round(b1 + 2 * b2 / g1 + 32 / g1 / g2, digits = 2))
+  } else {
+    return(round(as.numeric(sub("_", ".", config)), digits = 2))
+  }
+}
+
+load_data <- function(baseline_data_dir, mxq_data_dir, type) {
+  if (is.null(baseline_data_dir)) {
+    dat_dir <- path.expand(paste0("data/", type))
+    dat_fps <- dir(
+      path = dat_dir,
+      pattern = ".*\\.csv$",
+      recursive = TRUE,
+      full.names = TRUE
+    )
+  } else {
+    # Read csv files under baseline_data_dir, do not recursive into sub-folders
+    dat_dir <- path.expand(paste0(baseline_data_dir, "/", type))
+    dat_fps <- dir(
+      path = dat_dir,
+      pattern = ".*\\.csv$",
+      recursive = FALSE,
+      full.names = TRUE
+    )
+    mxq_dat_dir <- path.expand(paste0(mxq_data_dir, "/", type))
+    mxq_dat_fps <- dir(
+      path = mxq_dat_dir,
+      pattern = ".*\\.csv$",
+      recursive = TRUE,
+      full.names = TRUE
+    )
+    dat_fps <- c(dat_fps, mxq_dat_fps)
+  }
+  names(dat_fps) <- str_match(dat_fps, "mxq/(.*)/")[, 2]
+  df_combined <- ldply(
+    dat_fps,
+    read.csv,
+    stringsAsFactors = FALSE,
+    .id = "attempt"
+  )
+  return(df_combined)
+}
+
+parser <- OptionParser()
+parser <- add_option(
+  parser, c("-b", "--baseline_data_dir"),
+  type = "character",
+  help = "Data directory of baseline results",
+  metavar = "character"
+)
+parser <- add_option(
+  parser, c("-m", "--mxq_data_dir"),
+  type = "character",
+  help = "Data directory of MXQ results organized by attempt",
+  metavar = "character"
+)
+args <- parse_args(parser)
+
+if (is.null(args$baseline_data_dir)) {
+  df_combined_stor <- load_data(NULL, NULL, "stor")
+  df_combined_ppl <- load_data(NULL, NULL, "ppl")
+  df_combined_qnt <- load_data(NULL, NULL, "qnt")
+  df_combined_allot <- load_data(NULL, NULL, "allot")
+} else {
+  df_combined_stor <- load_data(
+    args$baseline_data_dir, args$mxq_data_dir, "stor"
+  )
+  df_combined_ppl <- load_data(
+    args$baseline_data_dir, args$mxq_data_dir, "ppl"
+  )
+  df_combined_qnt <- load_data(
+    args$baseline_data_dir, args$mxq_data_dir, "qnt"
+  )
+  df_combined_allot <- load_data(
+    args$baseline_data_dir, args$mxq_data_dir, "allot"
+  )
+}
+
+df_combined_stor <- df_combined_stor |>
+  select(
+    model, algo, config, attempt, load_mem_allot, model_storage_size
+  ) |>
+  mutate(
+    load_mem_allot = round(load_mem_allot / 1024^3, digits = 2),
+    model_storage_size = round(model_storage_size / 1024^3, digits = 2),
+  )
+
+df_combined_ppl <- df_combined_ppl |>
+  select(
+    model, algo, config, attempt,
+    ppl_wikitext, ppl_c4, quant_duration
+  )
+
+df_combined_qnt <- df_combined_qnt |>
+  select(
+    model, algo, config, attempt, quant_duration
+  )
+
+df_model_params <- tribble(
+  ~model, ~param_count,
+  "Llama-2-7b-hf", 7000000000,
+  "Llama-2-13b-hf", 13000000000,
+  "Meta-Llama-3-8B", 8000000000
+)
+
+df_combined <- df_combined_ppl |>
+  left_join(df_combined_qnt, join_by(model, algo, config, attempt)) |>
+  left_join(df_combined_stor, join_by(model, algo, config, attempt)) |>
+  mutate(
+    quant_duration =
+      ifelse(quant_duration.x > 0, quant_duration.x, quant_duration.y)
+  ) |>
+  left_join(df_model_params, join_by(model)) |>
+  mutate(
+    bpp = sapply(config, calc_bpp)
+  ) |>
+  select(!c("quant_duration.x", "quant_duration.y")) |>
+  distinct(model, algo, config, attempt, .keep_all = TRUE) |>
+  relocate(bpp, .after = config)
+
+write_csv(df_combined, "data/combined.csv", na = "")
+write.xlsx(
+  df_combined, "data/combined.xlsx",
+  overwrite = TRUE, asTable = TRUE
+)
+
+write_csv(df_combined_allot, "data/quant-cfg-allocation.csv", na = "")
+write.xlsx(
+  df_combined, "data/quant-cfg-allocation.xlsx",
+  overwrite = TRUE, asTable = TRUE
+)

lm-quant-toolkit/new_lm-quant-toolkit/lm-quant-toolkit/data-vis/data-vis.Rproj

@@ -0,0 +1,14 @@
+Version: 1.0
+ProjectId: 9f27daae-0012-43a6-b6a5-ec594e1aadd7
+
+RestoreWorkspace: Default
+SaveWorkspace: Default
+AlwaysSaveHistory: Default
+
+EnableCodeIndexing: Yes
+UseSpacesForTab: Yes
+NumSpacesForTab: 2
+Encoding: UTF-8
+
+RnwWeave: Sweave
+LaTeX: pdfLaTeX