Q4_0 scale selection using RMSE #835
Draft
+428
−91
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Very interesting analysis and data 😄 Btw, I've been thinking a little bit about how to determine the scale factor to minimize the RMS and I am fairly certain that there is a straightforward way to compute the optimum value without search. I don't have the formula yet - just a strong intuition lol |
I'd love to see that, but while the error function seems to be Here's the plot of the very first block in the 7B model (first input value = 9.888411e-05): (This is just the sum of squared errors, I didn't bother with the square root and scaling by QK.) |
|
So as mentioned in #397 (comment) I believe I have an RMSE-optimal but very slow implementation of the scaling search... And your implementation gets extremely close!: You posted:
"optimal":
That probably about as good as we can hope for. Full output for verification - if you get a lower RMSE for any layer I have a bug :) quantize-stats outputnote: source model is f16 testing 226 layers with max size 131072000 q4_0::layers.0.attention.wk.weight : rmse 0.00292301, maxerr 0.07012939, 95pct<0.0060, median<0.0018 q4_0::layers.0.attention.wo.weight : rmse 0.00100800, maxerr 0.03466797, 95pct<0.0020, median<0.0008 q4_0::layers.0.attention.wq.weight : rmse 0.00299672, maxerr 0.04935372, 95pct<0.0062, median<0.0016 q4_0::layers.0.attention.wv.weight : rmse 0.00110658, maxerr 0.01175963, 95pct<0.0022, median<0.0008 q4_0::layers.0.feed_forward.w1.weight : rmse 0.00137271, maxerr 0.06394231, 95pct<0.0026, median<0.0012 q4_0::layers.0.feed_forward.w2.weight : rmse 0.00167287, maxerr 0.05469465, 95pct<0.0030, median<0.0014 q4_0::layers.0.feed_forward.w3.weight : rmse 0.00132967, maxerr 0.01823677, 95pct<0.0024, median<0.0012 q4_0::layers.1.attention.wk.weight : rmse 0.00282859, maxerr 0.04061890, 95pct<0.0060, median<0.0016 q4_0::layers.1.attention.wo.weight : rmse 0.00097638, maxerr 0.03619385, 95pct<0.0020, median<0.0008 q4_0::layers.1.attention.wq.weight : rmse 0.00276326, maxerr 0.03581724, 95pct<0.0058, median<0.0016 q4_0::layers.1.attention.wv.weight : rmse 0.00094906, maxerr 0.00811867, 95pct<0.0020, median<0.0008 q4_0::layers.1.feed_forward.w1.weight : rmse 0.00174292, maxerr 0.03725962, 95pct<0.0032, median<0.0014 q4_0::layers.1.feed_forward.w2.weight : rmse 0.00171300, maxerr 0.05225383, 95pct<0.0032, median<0.0014 q4_0::layers.1.feed_forward.w3.weight : rmse 0.00165360, maxerr 0.02176375, 95pct<0.0030, median<0.0014 q4_0::layers.10.attention.wk.weight : rmse 0.00225472, maxerr 0.02979095, 95pct<0.0044, median<0.0016 q4_0::layers.10.attention.wo.weight : rmse 0.00143767, maxerr 0.04605806, 95pct<0.0026, median<0.0012 q4_0::layers.10.attention.wq.weight : rmse 0.00222988, maxerr 0.03421440, 95pct<0.0044, median<0.0016 q4_0::layers.10.attention.wv.weight : rmse 0.00144258, maxerr 0.01459024, 95pct<0.0028, median<0.0012 q4_0::layers.10.feed_forward.w1.weight : rmse 0.00183416, maxerr 0.02703372, 95pct<0.0034, median<0.0014 q4_0::layers.10.feed_forward.w2.weight : rmse 0.00174484, maxerr 0.04180530, 95pct<0.0032, median<0.0014 q4_0::layers.10.feed_forward.w3.weight : rmse 0.00177285, maxerr 0.02142334, 95pct<0.0032, median<0.0014 q4_0::layers.11.attention.wk.weight : rmse 0.00233274, maxerr 0.02713823, 95pct<0.0046, median<0.0018 q4_0::layers.11.attention.wo.weight : rmse 0.00150656, maxerr 0.03012497, 95pct<0.0028, median<0.0012 q4_0::layers.11.attention.wq.weight : rmse 0.00229496, maxerr 0.04412842, 95pct<0.0044, median<0.0018 q4_0::layers.11.attention.wv.weight : rmse 0.00151707, maxerr 0.02018456, 95pct<0.0028, median<0.0012 q4_0::layers.11.feed_forward.w1.weight : rmse 0.00182944, maxerr 0.02489303, 95pct<0.0034, median<0.0014 q4_0::layers.11.feed_forward.w2.weight : rmse 0.00175960, maxerr 0.05431067, 95pct<0.0032, median<0.0014 q4_0::layers.11.feed_forward.w3.weight : rmse 0.00178396, maxerr 0.02583313, 95pct<0.0032, median<0.0014 q4_0::layers.12.attention.wk.weight : rmse 0.00222511, maxerr 0.02594195, 95pct<0.0044, median<0.0016 q4_0::layers.12.attention.wo.weight : rmse 0.00147925, maxerr 0.02380922, 95pct<0.0028, median<0.0012 q4_0::layers.12.attention.wq.weight : rmse 0.00218941, maxerr 0.03447523, 95pct<0.0042, median<0.0016 q4_0::layers.12.attention.wv.weight : rmse 0.00146629, maxerr 0.01153804, 95pct<0.0028, median<0.0012 q4_0::layers.12.feed_forward.w1.weight : rmse 0.00183979, maxerr 0.03383104, 95pct<0.0034, median<0.0014 q4_0::layers.12.feed_forward.w2.weight : rmse 0.00176264, maxerr 0.05683154, 95pct<0.0032, median<0.0014 q4_0::layers.12.feed_forward.w3.weight : rmse 0.00179476, maxerr 0.01740211, 95pct<0.0034, median<0.0014 q4_0::layers.13.attention.wk.weight : rmse 0.00217331, maxerr 0.02676816, 95pct<0.0044, median<0.0016 q4_0::layers.13.attention.wo.weight : rmse 0.00153305, maxerr 0.04341370, 95pct<0.0028, median<0.0012 q4_0::layers.13.attention.wq.weight : rmse 0.00213820, maxerr 0.03543091, 95pct<0.0042, median<0.0016 q4_0::layers.13.attention.wv.weight : rmse 0.00153372, maxerr 0.01126552, 95pct<0.0028, median<0.0012 q4_0::layers.13.feed_forward.w1.weight : rmse 0.00183155, maxerr 0.02292150, 95pct<0.0034, median<0.0014 q4_0::layers.13.feed_forward.w2.weight : rmse 0.00177651, maxerr 0.03530073, 95pct<0.0032, median<0.0014 q4_0::layers.13.feed_forward.w3.weight : rmse 0.00181181, maxerr 0.01798833, 95pct<0.0034, median<0.0016 q4_0::layers.14.attention.wk.weight : rmse 0.00217185, maxerr 0.02497105, 95pct<0.0042, median<0.0016 q4_0::layers.14.attention.wo.weight : rmse 0.00153627, maxerr 0.06212232, 95pct<0.0028, median<0.0012 q4_0::layers.14.attention.wq.weight : rmse 0.00215347, maxerr 0.03887939, 95pct<0.0042, median<0.0016 q4_0::layers.14.attention.wv.weight : rmse 0.00154264, maxerr 0.01345214, 95pct<0.0028, median<0.0012 q4_0::layers.14.feed_forward.w1.weight : rmse 0.00182898, maxerr 0.02304077, 95pct<0.0034, median<0.0014 q4_0::layers.14.feed_forward.w2.weight : rmse 0.00178511, maxerr 0.05890521, 95pct<0.0032, median<0.0014 q4_0::layers.14.feed_forward.w3.weight : rmse 0.00181856, maxerr 0.02665675, 95pct<0.0034, median<0.0016 q4_0::layers.15.attention.wk.weight : rmse 0.00219269, maxerr 0.02394998, 95pct<0.0044, median<0.0016 q4_0::layers.15.attention.wo.weight : rmse 0.00154000, maxerr 0.02813050, 95pct<0.0028, median<0.0012 q4_0::layers.15.attention.wq.weight : rmse 0.00215290, maxerr 0.03628540, 95pct<0.0042, median<0.0016 q4_0::layers.15.attention.wv.weight : rmse 0.00154661, maxerr 0.01409675, 95pct<0.0028, median<0.0012 q4_0::layers.15.feed_forward.w1.weight : rmse 0.00182940, maxerr 0.02419187, 95pct<0.0034, median<0.0014 q4_0::layers.15.feed_forward.w2.weight : rmse 0.00178558, maxerr 0.05858561, 95pct<0.0032, median<0.0014 q4_0::layers.15.feed_forward.w3.weight : rmse 0.00181912, maxerr 0.02241516, 95pct<0.0034, median<0.0016 q4_0::layers.16.attention.wk.weight : rmse 0.00217754, maxerr 0.02458954, 95pct<0.0042, median<0.0016 q4_0::layers.16.attention.wo.weight : rmse 0.00163187, maxerr 0.05107081, 95pct<0.0030, median<0.0014 q4_0::layers.16.attention.wq.weight : rmse 0.00212385, maxerr 0.04119629, 95pct<0.0040, median<0.0016 q4_0::layers.16.attention.wv.weight : rmse 0.00164553, maxerr 0.01337417, 95pct<0.0030, median<0.0014 q4_0::layers.16.feed_forward.w1.weight : rmse 0.00184241, maxerr 0.02344798, 95pct<0.0034, median<0.0016 q4_0::layers.16.feed_forward.w2.weight : rmse 0.00178439, maxerr 0.05552104, 95pct<0.0032, median<0.0014 q4_0::layers.16.feed_forward.w3.weight : rmse 0.00181314, maxerr 0.02277143, 95pct<0.0034, median<0.0016 q4_0::layers.17.attention.wk.weight : rmse 0.00212176, maxerr 0.02422421, 95pct<0.0042, median<0.0016 q4_0::layers.17.attention.wo.weight : rmse 0.00165387, maxerr 0.03002930, 95pct<0.0030, median<0.0014 q4_0::layers.17.attention.wq.weight : rmse 0.00207895, maxerr 0.04604350, 95pct<0.0040, median<0.0016 q4_0::layers.17.attention.wv.weight : rmse 0.00165649, maxerr 0.01419830, 95pct<0.0030, median<0.0014 q4_0::layers.17.feed_forward.w1.weight : rmse 0.00184599, maxerr 0.02392328, 95pct<0.0034, median<0.0016 q4_0::layers.17.feed_forward.w2.weight : rmse 0.00179142, maxerr 0.04622682, 95pct<0.0032, median<0.0014 q4_0::layers.17.feed_forward.w3.weight : rmse 0.00181806, maxerr 0.02359099, 95pct<0.0034, median<0.0016 q4_0::layers.18.attention.wk.weight : rmse 0.00208260, maxerr 0.02502441, 95pct<0.0040, median<0.0016 q4_0::layers.18.attention.wo.weight : rmse 0.00164773, maxerr 0.03822631, 95pct<0.0030, median<0.0014 q4_0::layers.18.attention.wq.weight : rmse 0.00205646, maxerr 0.04051746, 95pct<0.0040, median<0.0016 q4_0::layers.18.attention.wv.weight : rmse 0.00165172, maxerr 0.01335841, 95pct<0.0030, median<0.0014 q4_0::layers.18.feed_forward.w1.weight : rmse 0.00186100, maxerr 0.03084695, 95pct<0.0034, median<0.0016 q4_0::layers.18.feed_forward.w2.weight : rmse 0.00178702, maxerr 0.06258377, 95pct<0.0032, median<0.0014 q4_0::layers.18.feed_forward.w3.weight : rmse 0.00181154, maxerr 0.01813507, 95pct<0.0034, median<0.0016 q4_0::layers.19.attention.wk.weight : rmse 0.00204409, maxerr 0.02549587, 95pct<0.0040, median<0.0016 q4_0::layers.19.attention.wo.weight : rmse 0.00171742, maxerr 0.04106662, 95pct<0.0032, median<0.0014 q4_0::layers.19.attention.wq.weight : rmse 0.00202074, maxerr 0.04685394, 95pct<0.0040, median<0.0016 q4_0::layers.19.attention.wv.weight : rmse 0.00173205, maxerr 0.01311102, 95pct<0.0032, median<0.0014 q4_0::layers.19.feed_forward.w1.weight : rmse 0.00187151, maxerr 0.03121948, 95pct<0.0034, median<0.0016 q4_0::layers.19.feed_forward.w2.weight : rmse 0.00178935, maxerr 0.04564381, 95pct<0.0032, median<0.0014 q4_0::layers.19.feed_forward.w3.weight : rmse 0.00180759, maxerr 0.02138457, 95pct<0.0034, median<0.0016 q4_0::layers.2.attention.wk.weight : rmse 0.00310555, maxerr 0.03675859, 95pct<0.0064, median<0.0020 q4_0::layers.2.attention.wo.weight : rmse 0.00115159, maxerr 0.04546779, 95pct<0.0022, median<0.0010 q4_0::layers.2.attention.wq.weight : rmse 0.00298841, maxerr 0.03752440, 95pct<0.0060, median<0.0020 q4_0::layers.2.attention.wv.weight : rmse 0.00112951, maxerr 0.00926531, 95pct<0.0022, median<0.0010 q4_0::layers.2.feed_forward.w1.weight : rmse 0.00183671, maxerr 0.05353853, 95pct<0.0034, median<0.0016 q4_0::layers.2.feed_forward.w2.weight : rmse 0.00170433, maxerr 0.09649658, 95pct<0.0032, median<0.0014 q4_0::layers.2.feed_forward.w3.weight : rmse 0.00167454, maxerr 0.03201294, 95pct<0.0030, median<0.0014 q4_0::layers.20.attention.wk.weight : rmse 0.00207524, maxerr 0.02473852, 95pct<0.0040, median<0.0016 q4_0::layers.20.attention.wo.weight : rmse 0.00176106, maxerr 0.02588722, 95pct<0.0032, median<0.0014 q4_0::layers.20.attention.wq.weight : rmse 0.00204837, maxerr 0.05462646, 95pct<0.0040, median<0.0016 q4_0::layers.20.attention.wv.weight : rmse 0.00178526, maxerr 0.01499712, 95pct<0.0034, median<0.0014 q4_0::layers.20.feed_forward.w1.weight : rmse 0.00188099, maxerr 0.02917725, 95pct<0.0034, median<0.0016 q4_0::layers.20.feed_forward.w2.weight : rmse 0.00179125, maxerr 0.06890869, 95pct<0.0032, median<0.0014 q4_0::layers.20.feed_forward.w3.weight : rmse 0.00180859, maxerr 0.01596069, 95pct<0.0034, median<0.0016 q4_0::layers.21.attention.wk.weight : rmse 0.00200054, maxerr 0.02908368, 95pct<0.0040, median<0.0014 q4_0::layers.21.attention.wo.weight : rmse 0.00177119, maxerr 0.05007464, 95pct<0.0032, median<0.0014 q4_0::layers.21.attention.wq.weight : rmse 0.00198177, maxerr 0.05149466, 95pct<0.0038, median<0.0014 q4_0::layers.21.attention.wv.weight : rmse 0.00179837, maxerr 0.01333202, 95pct<0.0034, median<0.0014 q4_0::layers.21.feed_forward.w1.weight : rmse 0.00189033, maxerr 0.03076535, 95pct<0.0034, median<0.0016 q4_0::layers.21.feed_forward.w2.weight : rmse 0.00178966, maxerr 0.03637502, 95pct<0.0032, median<0.0014 q4_0::layers.21.feed_forward.w3.weight : rmse 0.00180614, maxerr 0.02140096, 95pct<0.0034, median<0.0016 q4_0::layers.22.attention.wk.weight : rmse 0.00203025, maxerr 0.03339660, 95pct<0.0040, median<0.0016 q4_0::layers.22.attention.wo.weight : rmse 0.00177702, maxerr 0.07931513, 95pct<0.0032, median<0.0014 q4_0::layers.22.attention.wq.weight : rmse 0.00201616, maxerr 0.04454328, 95pct<0.0038, median<0.0016 q4_0::layers.22.attention.wv.weight : rmse 0.00178748, maxerr 0.01423188, 95pct<0.0034, median<0.0014 q4_0::layers.22.feed_forward.w1.weight : rmse 0.00189302, maxerr 0.02517700, 95pct<0.0034, median<0.0016 q4_0::layers.22.feed_forward.w2.weight : rmse 0.00179775, maxerr 0.04281616, 95pct<0.0034, median<0.0016 q4_0::layers.22.feed_forward.w3.weight : rmse 0.00181394, maxerr 0.03024019, 95pct<0.0034, median<0.0016 q4_0::layers.23.attention.wk.weight : rmse 0.00195991, maxerr 0.02972737, 95pct<0.0038, median<0.0014 q4_0::layers.23.attention.wo.weight : rmse 0.00182629, maxerr 0.04887556, 95pct<0.0034, median<0.0016 q4_0::layers.23.attention.wq.weight : rmse 0.00195417, maxerr 0.04232788, 95pct<0.0038, median<0.0014 q4_0::layers.23.attention.wv.weight : rmse 0.00185857, maxerr 0.01577342, 95pct<0.0034, median<0.0016 q4_0::layers.23.feed_forward.w1.weight : rmse 0.00189658, maxerr 0.03308105, 95pct<0.0034, median<0.0016 q4_0::layers.23.feed_forward.w2.weight : rmse 0.00180367, maxerr 0.04928589, 95pct<0.0034, median<0.0016 q4_0::layers.23.feed_forward.w3.weight : rmse 0.00181737, maxerr 0.02468872, 95pct<0.0034, median<0.0016 q4_0::layers.24.attention.wk.weight : rmse 0.00196715, maxerr 0.02162942, 95pct<0.0038, median<0.0014 q4_0::layers.24.attention.wo.weight : rmse 0.00184930, maxerr 0.03620195, 95pct<0.0034, median<0.0016 q4_0::layers.24.attention.wq.weight : rmse 0.00195618, maxerr 0.04705903, 95pct<0.0038, median<0.0014 q4_0::layers.24.attention.wv.weight : rmse 0.00188009, maxerr 0.01770980, 95pct<0.0034, median<0.0016 q4_0::layers.24.feed_forward.w1.weight : rmse 0.00189906, maxerr 0.02117351, 95pct<0.0034, median<0.0016 q4_0::layers.24.feed_forward.w2.weight : rmse 0.00181186, maxerr 0.05899048, 95pct<0.0034, median<0.0016 q4_0::layers.24.feed_forward.w3.weight : rmse 0.00182756, maxerr 0.02068704, 95pct<0.0034, median<0.0016 q4_0::layers.25.attention.wk.weight : rmse 0.00202900, maxerr 0.02362627, 95pct<0.0038, median<0.0016 q4_0::layers.25.attention.wo.weight : rmse 0.00186576, maxerr 0.06477863, 95pct<0.0034, median<0.0016 q4_0::layers.25.attention.wq.weight : rmse 0.00200834, maxerr 0.03808594, 95pct<0.0038, median<0.0016 q4_0::layers.25.attention.wv.weight : rmse 0.00188682, maxerr 0.01595676, 95pct<0.0034, median<0.0016 q4_0::layers.25.feed_forward.w1.weight : rmse 0.00190352, maxerr 0.02079988, 95pct<0.0036, median<0.0016 q4_0::layers.25.feed_forward.w2.weight : rmse 0.00181823, maxerr 0.03286743, 95pct<0.0034, median<0.0016 q4_0::layers.25.feed_forward.w3.weight : rmse 0.00183412, maxerr 0.01735053, 95pct<0.0034, median<0.0016 q4_0::layers.26.attention.wk.weight : rmse 0.00199544, maxerr 0.02733952, 95pct<0.0038, median<0.0016 q4_0::layers.26.attention.wo.weight : rmse 0.00192010, maxerr 0.02563220, 95pct<0.0036, median<0.0016 q4_0::layers.26.attention.wq.weight : rmse 0.00197604, maxerr 0.03735352, 95pct<0.0038, median<0.0016 q4_0::layers.26.attention.wv.weight : rmse 0.00194300, maxerr 0.01509885, 95pct<0.0036, median<0.0016 q4_0::layers.26.feed_forward.w1.weight : rmse 0.00190232, maxerr 0.03396144, 95pct<0.0036, median<0.0016 q4_0::layers.26.feed_forward.w2.weight : rmse 0.00183005, maxerr 0.04354858, 95pct<0.0034, median<0.0016 q4_0::layers.26.feed_forward.w3.weight : rmse 0.00184771, maxerr 0.03059387, 95pct<0.0034, median<0.0016 q4_0::layers.27.attention.wk.weight : rmse 0.00198943, maxerr 0.02681477, 95pct<0.0038, median<0.0016 q4_0::layers.27.attention.wo.weight : rmse 0.00196662, maxerr 0.05517289, 95pct<0.0036, median<0.0016 q4_0::layers.27.attention.wq.weight : rmse 0.00198182, maxerr 0.03899045, 95pct<0.0038, median<0.0016 q4_0::layers.27.attention.wv.weight : rmse 0.00197615, maxerr 0.01669417, 95pct<0.0036, median<0.0016 q4_0::layers.27.feed_forward.w1.weight : rmse 0.00190184, maxerr 0.02731323, 95pct<0.0036, median<0.0016 q4_0::layers.27.feed_forward.w2.weight : rmse 0.00184141, maxerr 0.04620361, 95pct<0.0034, median<0.0016 q4_0::layers.27.feed_forward.w3.weight : rmse 0.00185554, maxerr 0.04153442, 95pct<0.0034, median<0.0016 q4_0::layers.28.attention.wk.weight : rmse 0.00194652, maxerr 0.02850908, 95pct<0.0038, median<0.0014 q4_0::layers.28.attention.wo.weight : rmse 0.00198808, maxerr 0.03118306, 95pct<0.0036, median<0.0016 q4_0::layers.28.attention.wq.weight : rmse 0.00194074, maxerr 0.04092407, 95pct<0.0038, median<0.0014 q4_0::layers.28.attention.wv.weight : rmse 0.00198781, maxerr 0.01527815, 95pct<0.0036, median<0.0016 q4_0::layers.28.feed_forward.w1.weight : rmse 0.00189349, maxerr 0.03170776, 95pct<0.0034, median<0.0016 q4_0::layers.28.feed_forward.w2.weight : rmse 0.00185116, maxerr 0.05222714, 95pct<0.0034, median<0.0016 q4_0::layers.28.feed_forward.w3.weight : rmse 0.00186482, maxerr 0.03631857, 95pct<0.0034, median<0.0016 q4_0::layers.29.attention.wk.weight : rmse 0.00193214, maxerr 0.02202798, 95pct<0.0038, median<0.0014 q4_0::layers.29.attention.wo.weight : rmse 0.00204716, maxerr 0.03959709, 95pct<0.0038, median<0.0016 q4_0::layers.29.attention.wq.weight : rmse 0.00192283, maxerr 0.04244995, 95pct<0.0036, median<0.0014 q4_0::layers.29.attention.wv.weight : rmse 0.00204643, maxerr 0.01519237, 95pct<0.0038, median<0.0016 q4_0::layers.29.feed_forward.w1.weight : rmse 0.00189820, maxerr 0.03314209, 95pct<0.0036, median<0.0016 q4_0::layers.29.feed_forward.w2.weight : rmse 0.00186130, maxerr 0.09802246, 95pct<0.0034, median<0.0016 q4_0::layers.29.feed_forward.w3.weight : rmse 0.00187583, maxerr 0.02655141, 95pct<0.0034, median<0.0016 q4_0::layers.3.attention.wk.weight : rmse 0.00257589, maxerr 0.03777078, 95pct<0.0052, median<0.0018 q4_0::layers.3.attention.wo.weight : rmse 0.00133662, maxerr 0.04435936, 95pct<0.0024, median<0.0012 q4_0::layers.3.attention.wq.weight : rmse 0.00246442, maxerr 0.04611765, 95pct<0.0048, median<0.0018 q4_0::layers.3.attention.wv.weight : rmse 0.00133929, maxerr 0.01030663, 95pct<0.0024, median<0.0012 q4_0::layers.3.feed_forward.w1.weight : rmse 0.00185664, maxerr 0.03087639, 95pct<0.0034, median<0.0016 q4_0::layers.3.feed_forward.w2.weight : rmse 0.00171196, maxerr 0.05057278, 95pct<0.0032, median<0.0014 q4_0::layers.3.feed_forward.w3.weight : rmse 0.00170679, maxerr 0.02278137, 95pct<0.0032, median<0.0014 q4_0::layers.30.attention.wk.weight : rmse 0.00195269, maxerr 0.03295821, 95pct<0.0038, median<0.0016 q4_0::layers.30.attention.wo.weight : rmse 0.00204545, maxerr 0.05445015, 95pct<0.0038, median<0.0016 q4_0::layers.30.attention.wq.weight : rmse 0.00194719, maxerr 0.04063878, 95pct<0.0036, median<0.0016 q4_0::layers.30.attention.wv.weight : rmse 0.00202005, maxerr 0.01512921, 95pct<0.0038, median<0.0016 q4_0::layers.30.feed_forward.w1.weight : rmse 0.00191074, maxerr 0.02958679, 95pct<0.0036, median<0.0016 q4_0::layers.30.feed_forward.w2.weight : rmse 0.00191046, maxerr 0.14257812, 95pct<0.0034, median<0.0016 q4_0::layers.30.feed_forward.w3.weight : rmse 0.00189492, maxerr 0.04852676, 95pct<0.0034, median<0.0016 q4_0::layers.31.attention.wk.weight : rmse 0.00201812, maxerr 0.02451627, 95pct<0.0038, median<0.0016 q4_0::layers.31.attention.wo.weight : rmse 0.00184503, maxerr 0.11907780, 95pct<0.0034, median<0.0014 q4_0::layers.31.attention.wq.weight : rmse 0.00197563, maxerr 0.02724165, 95pct<0.0038, median<0.0016 q4_0::layers.31.attention.wv.weight : rmse 0.00182399, maxerr 0.01841706, 95pct<0.0034, median<0.0014 q4_0::layers.31.feed_forward.w1.weight : rmse 0.00199676, maxerr 0.03135899, 95pct<0.0036, median<0.0016 q4_0::layers.31.feed_forward.w2.weight : rmse 0.00191905, maxerr 0.11260986, 95pct<0.0036, median<0.0016 q4_0::layers.31.feed_forward.w3.weight : rmse 0.00197545, maxerr 0.04486084, 95pct<0.0036, median<0.0016 q4_0::layers.4.attention.wk.weight : rmse 0.00252572, maxerr 0.03471547, 95pct<0.0050, median<0.0018 q4_0::layers.4.attention.wo.weight : rmse 0.00133709, maxerr 0.05675527, 95pct<0.0026, median<0.0012 q4_0::layers.4.attention.wq.weight : rmse 0.00250660, maxerr 0.04748535, 95pct<0.0048, median<0.0018 q4_0::layers.4.attention.wv.weight : rmse 0.00133764, maxerr 0.01021584, 95pct<0.0026, median<0.0012 q4_0::layers.4.feed_forward.w1.weight : rmse 0.00188008, maxerr 0.03756605, 95pct<0.0034, median<0.0016 q4_0::layers.4.feed_forward.w2.weight : rmse 0.00170612, maxerr 0.04783656, 95pct<0.0032, median<0.0014 q4_0::layers.4.feed_forward.w3.weight : rmse 0.00171322, maxerr 0.03393555, 95pct<0.0032, median<0.0014 q4_0::layers.5.attention.wk.weight : rmse 0.00238210, maxerr 0.03174898, 95pct<0.0046, median<0.0018 q4_0::layers.5.attention.wo.weight : rmse 0.00135344, maxerr 0.04260254, 95pct<0.0026, median<0.0012 q4_0::layers.5.attention.wq.weight : rmse 0.00236603, maxerr 0.04248789, 95pct<0.0046, median<0.0018 q4_0::layers.5.attention.wv.weight : rmse 0.00136147, maxerr 0.01390839, 95pct<0.0026, median<0.0012 q4_0::layers.5.feed_forward.w1.weight : rmse 0.00191865, maxerr 0.03069225, 95pct<0.0036, median<0.0016 q4_0::layers.5.feed_forward.w2.weight : rmse 0.00168901, maxerr 0.04306030, 95pct<0.0032, median<0.0014 q4_0::layers.5.feed_forward.w3.weight : rmse 0.00170621, maxerr 0.02728271, 95pct<0.0032, median<0.0014 q4_0::layers.6.attention.wk.weight : rmse 0.00243652, maxerr 0.02662471, 95pct<0.0048, median<0.0018 q4_0::layers.6.attention.wo.weight : rmse 0.00136724, maxerr 0.06586111, 95pct<0.0026, median<0.0012 q4_0::layers.6.attention.wq.weight : rmse 0.00238362, maxerr 0.04891968, 95pct<0.0046, median<0.0018 q4_0::layers.6.attention.wv.weight : rmse 0.00137151, maxerr 0.01011706, 95pct<0.0026, median<0.0012 q4_0::layers.6.feed_forward.w1.weight : rmse 0.00189249, maxerr 0.04025269, 95pct<0.0036, median<0.0016 q4_0::layers.6.feed_forward.w2.weight : rmse 0.00170650, maxerr 0.04733700, 95pct<0.0032, median<0.0014 q4_0::layers.6.feed_forward.w3.weight : rmse 0.00172804, maxerr 0.02388000, 95pct<0.0032, median<0.0014 q4_0::layers.7.attention.wk.weight : rmse 0.00236151, maxerr 0.02808842, 95pct<0.0046, median<0.0018 q4_0::layers.7.attention.wo.weight : rmse 0.00139671, maxerr 0.03219414, 95pct<0.0026, median<0.0012 q4_0::layers.7.attention.wq.weight : rmse 0.00234200, maxerr 0.04630763, 95pct<0.0046, median<0.0018 q4_0::layers.7.attention.wv.weight : rmse 0.00141431, maxerr 0.01132994, 95pct<0.0026, median<0.0012 q4_0::layers.7.feed_forward.w1.weight : rmse 0.00187467, maxerr 0.03225514, 95pct<0.0034, median<0.0016 q4_0::layers.7.feed_forward.w2.weight : rmse 0.00171376, maxerr 0.04287861, 95pct<0.0032, median<0.0014 q4_0::layers.7.feed_forward.w3.weight : rmse 0.00173478, maxerr 0.02548218, 95pct<0.0032, median<0.0014 q4_0::layers.8.attention.wk.weight : rmse 0.00230324, maxerr 0.02782059, 95pct<0.0046, median<0.0016 q4_0::layers.8.attention.wo.weight : rmse 0.00139008, maxerr 0.03478485, 95pct<0.0026, median<0.0012 q4_0::layers.8.attention.wq.weight : rmse 0.00230350, maxerr 0.04038759, 95pct<0.0046, median<0.0016 q4_0::layers.8.attention.wv.weight : rmse 0.00140037, maxerr 0.01309800, 95pct<0.0026, median<0.0012 q4_0::layers.8.feed_forward.w1.weight : rmse 0.00187512, maxerr 0.03340167, 95pct<0.0034, median<0.0016 q4_0::layers.8.feed_forward.w2.weight : rmse 0.00171484, maxerr 0.03771973, 95pct<0.0032, median<0.0014 q4_0::layers.8.feed_forward.w3.weight : rmse 0.00173956, maxerr 0.02236661, 95pct<0.0032, median<0.0014 q4_0::layers.9.attention.wk.weight : rmse 0.00224144, maxerr 0.02832547, 95pct<0.0044, median<0.0016 q4_0::layers.9.attention.wo.weight : rmse 0.00137911, maxerr 0.03645405, 95pct<0.0026, median<0.0012 q4_0::layers.9.attention.wq.weight : rmse 0.00222848, maxerr 0.04025269, 95pct<0.0044, median<0.0016 q4_0::layers.9.attention.wv.weight : rmse 0.00139053, maxerr 0.01049893, 95pct<0.0026, median<0.0012 q4_0::layers.9.feed_forward.w1.weight : rmse 0.00184797, maxerr 0.04046337, 95pct<0.0034, median<0.0016 q4_0::layers.9.feed_forward.w2.weight : rmse 0.00172856, maxerr 0.04580688, 95pct<0.0032, median<0.0014 q4_0::layers.9.feed_forward.w3.weight : rmse 0.00175109, maxerr 0.04849243, 95pct<0.0032, median<0.0014 q4_0::output.weight : rmse 0.00165429, maxerr 0.02467346, 95pct<0.0032, median<0.0014 q4_0::tok_embeddings.weight : rmse 0.00163455, maxerr 0.01590976, 95pct<0.0030, median<0.0014 q4_0 : rmse 0.00184913, maxerr 0.14257812, 95pct<0.0034, median<0.0014 |
|
Now that the statistics tool has landed in master, I've rebased my branch and updated the tool to accept an @unbounded : I will definitively have a look at your approach, thanks a lot. edit: pulled in your commit and updated the stats tool. It is indeed slow ;-). 80% of time is spent in
|
Use a sweep line approach to scan all configurations of quantization, examining every changeover point where a quantize value changes, and find the optimal scaling for each configuration analytically.
| @@ -1,7 +1,11 @@ | |||
| 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d models/7B/consolidated.00.pth | |||
| 0cc0b0a3dc8cd29f005946f8364ac2bbce797e792a40c0fb4114615e4f825976 models/7B/ggml-model-f16.bin | |||
| 5dec1979849d73e361a8bcc10bc8f53237cbbe435a572882dc87629e011e24b3 models/7B/ggml-model-q4_0.bin | |||
There was a problem hiding this comment.
Could you please remove quantized models, since everyone would have their unique quantized models.
There was a problem hiding this comment.
The idea would be that model generation is deterministic across platforms and SIMD optimizations, so the files should be identical. Of course if you keep your Q4_0 files without updating to minor version 1, this wouldn't match. I might remove it for this PR, but in the long term I think it's a good idea to ensure everyone uses the same inputs.
There was a problem hiding this comment.
OK, I have generated new quantized model and checksum mathes with yours.
There was a problem hiding this comment.
I am sorry, is this checksum is for q4_0 that have no minor version yet?
Edit: Oh, I see, for minor v1. 4 bytes long than previous version 😅
| @@ -644,7 +644,7 @@ static bool llama_model_load( | |||
| size_t total_size = 0; | |||
| model.n_loaded = 0; | |||
|
|
|||
| while (true) { | |||
| while (size_t(fin.tellg()) + 12 < file_size) { | |||
There was a problem hiding this comment.
I'd rather do
int offset = 0;
...
offset += sizeof(total_size) + sizeof(model.n_loaded)
There was a problem hiding this comment.
total_size and model.n_loaded are not written or read from the file, so I don't understand why you would use their sizeof.
I admit that the + 12 could be written better. It is intended to be sizeof(n_dims) + sizeof(length) + sizeof(ftype), the next three elements being read.
| } else { | ||
| fprintf(stderr, "error: %s not in list of implementations\n", argv[i]); | ||
| invalid_param = true; | ||
| } | ||
| } else if (arg == "-v") { |
There was a problem hiding this comment.
Could you please add || --verbose?
|
Initial perplexity test. q4_0, MINOR 0, w/o BLAS: Commit 678e138 (shown as Final score
And closer: |
|
Leaving another comment to let you know final perplexity [655]6.5655. Perplexity discussion for previous results. |
|
@ivanstepanovftw Thanks for your effort. The first few values match mine exactly, so I'll trust your results. It's good to see at least a small improvement. But as I said in #397, maybe the RMSE of the quantization is a distraction. This method leads to a mean scale value of 8.092, so there will be clipping of the maximum value. I would like to see us experiment with #729 but with more (larger) scale values instead of just 7 or 8. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This combines some ideas from PR #729 and issue #397 to select a scale factor for Q4_0 with low RMS error.
In order to KISS, I simply made a table of 8 hard-coded values, after analysing the optimum values in steps of 0.1.
The result of that analysis is documented in
examples/quantize/scale.pyand reproduced here:Error statistics (#728):
quantize.cpp run time on 7B:
I introduce a minor version number at the very end of the file.
This allows us to nudge the user to re-create their files without breaking anything.
I had to modify the read loop, as it used to try to read past EOF.
I removed the test of
ggml_quantize_q4_0which I originally wrote and was quite minimal.This is admittedly lazy, but I couldn't think of a good test right away.
Maybe we just need to provide a model file that's not too big for the CI machines and check for equivalence after quantization.
The alignment macros are a bit of a hack. I don't have Windows to test here and don't want to keep hitting the CI with trial-and-error.
Is there a clean cross-platform way to do it? And come to think about alignment, why are the input
floats not aligned? (edit: probably becausellama_model_quantize_internaldoesn't usemmap, let me see if we can force the alignment of the buffers).Currently running perplexity, but it's taking 12 hours here so I may not wait for that.
This does not obsolete #729, as my PR only changes the method for the model generation.
We might still use @unbounded's work and set the scale to -8 instead of +7 for the other uses of the quantization function.