Remove stale KLD log from the previous patch
Browse files- benchmarks/kld_final.log +0 -80
benchmarks/kld_final.log
DELETED
|
@@ -1,80 +0,0 @@
|
|
| 1 |
-
0.01.766.660 I cmn init: llama threadpool init, n_threads = 16
|
| 2 |
-
0.01.845.751 I
|
| 3 |
-
0.01.845.855 I system_info: n_threads = 16 (n_threads_batch = 16) / 192 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VBMI = 1 | AVX512_VNNI = 1 | AVX512_BF16 = 1 | AMX_INT8 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
|
| 4 |
-
0.01.886.689 I kl_divergence: computing over 32 chunks, n_ctx=512, batch_size=512, n_seq=1
|
| 5 |
-
0.04.588.983 I kl_divergence: 2.70 seconds per pass - ETA 1.43 minutes
|
| 6 |
-
|
| 7 |
-
chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p
|
| 8 |
-
1 31.8950 ± 10.7913 -0.01919 ± 0.12414 0.47995 ± 0.07714 16.215 ± 2.303 % 78.824 ± 2.564 %
|
| 9 |
-
2 18.7693 ± 3.8844 0.00294 ± 0.06270 0.25980 ± 0.04020 12.069 ± 1.576 % 85.294 ± 1.570 %
|
| 10 |
-
3 15.4713 ± 2.4439 0.01637 ± 0.04302 0.21289 ± 0.03283 11.207 ± 1.271 % 87.712 ± 1.188 %
|
| 11 |
-
4 14.1940 ± 1.9658 0.03541 ± 0.03515 0.22116 ± 0.02998 10.684 ± 1.035 % 88.431 ± 1.002 %
|
| 12 |
-
5 10.8930 ± 1.3155 0.01565 ± 0.03140 0.23288 ± 0.02782 11.670 ± 0.962 % 89.020 ± 0.876 %
|
| 13 |
-
6 7.3329 ± 0.7564 0.01319 ± 0.02617 0.19435 ± 0.02328 10.668 ± 0.879 % 90.850 ± 0.737 %
|
| 14 |
-
7 5.5187 ± 0.4963 0.01079 ± 0.02243 0.16887 ± 0.02007 9.970 ± 0.809 % 92.157 ± 0.637 %
|
| 15 |
-
8 4.4589 ± 0.3553 0.00925 ± 0.01963 0.14834 ± 0.01760 9.372 ± 0.755 % 93.088 ± 0.562 %
|
| 16 |
-
9 4.0899 ± 0.2965 0.00299 ± 0.01793 0.14117 ± 0.01709 9.272 ± 0.730 % 93.595 ± 0.511 %
|
| 17 |
-
10 4.9979 ± 0.3706 0.01408 ± 0.02749 0.25037 ± 0.02457 10.943 ± 0.721 % 91.647 ± 0.548 %
|
| 18 |
-
11 4.5721 ± 0.3136 0.01320 ± 0.02500 0.22823 ± 0.02238 10.458 ± 0.686 % 92.228 ± 0.506 %
|
| 19 |
-
12 4.1937 ± 0.2670 0.01381 ± 0.02296 0.21160 ± 0.02057 10.275 ± 0.655 % 92.647 ± 0.472 %
|
| 20 |
-
13 3.9080 ± 0.2322 0.01163 ± 0.02146 0.20177 ± 0.01926 10.305 ± 0.634 % 93.032 ± 0.442 %
|
| 21 |
-
14 3.7166 ± 0.2082 0.01202 ± 0.01996 0.18824 ± 0.01791 10.031 ± 0.610 % 93.389 ± 0.416 %
|
| 22 |
-
15 3.5969 ± 0.1918 0.01192 ± 0.01865 0.17664 ± 0.01673 9.818 ± 0.585 % 93.725 ± 0.392 %
|
| 23 |
-
16 3.6195 ± 0.1845 0.00955 ± 0.01752 0.16778 ± 0.01571 9.655 ± 0.562 % 93.775 ± 0.378 %
|
| 24 |
-
17 3.7784 ± 0.1872 0.00322 ± 0.01773 0.19185 ± 0.01589 10.467 ± 0.555 % 93.218 ± 0.382 %
|
| 25 |
-
18 3.9020 ± 0.1906 0.01415 ± 0.01814 0.22527 ± 0.01680 11.350 ± 0.545 % 92.702 ± 0.384 %
|
| 26 |
-
19 3.9563 ± 0.1872 0.01320 ± 0.01723 0.21543 ± 0.01595 11.213 ± 0.531 % 92.859 ± 0.370 %
|
| 27 |
-
20 4.0206 ± 0.1852 0.00258 ± 0.01733 0.22077 ± 0.01576 11.437 ± 0.520 % 92.647 ± 0.366 %
|
| 28 |
-
21 4.2279 ± 0.1948 0.01021 ± 0.01782 0.25106 ± 0.01716 12.200 ± 0.521 % 92.176 ± 0.367 %
|
| 29 |
-
22 4.2137 ± 0.1895 0.01000 ± 0.01712 0.24988 ± 0.01667 12.241 ± 0.508 % 92.103 ± 0.360 %
|
| 30 |
-
23 4.1466 ± 0.1817 0.01604 ± 0.01705 0.24787 ± 0.01625 12.207 ± 0.495 % 92.242 ± 0.349 %
|
| 31 |
-
24 4.1800 ± 0.1779 0.01373 ± 0.01640 0.23897 ± 0.01559 11.969 ± 0.484 % 92.435 ± 0.338 %
|
| 32 |
-
25 4.1542 ± 0.1722 0.01618 ± 0.01584 0.23559 ± 0.01505 12.003 ± 0.474 % 92.502 ± 0.330 %
|
| 33 |
-
26 4.0994 ± 0.1649 0.01558 ± 0.01527 0.22948 ± 0.01454 11.892 ± 0.466 % 92.609 ± 0.321 %
|
| 34 |
-
27 4.1120 ± 0.1618 0.01565 ± 0.01479 0.22470 ± 0.01405 11.881 ± 0.458 % 92.694 ± 0.314 %
|
| 35 |
-
28 4.2240 ± 0.1640 0.01531 ± 0.01436 0.22426 ± 0.01375 11.893 ± 0.449 % 92.591 ± 0.310 %
|
| 36 |
-
29 4.2403 ± 0.1614 0.01391 ± 0.01395 0.22072 ± 0.01336 11.775 ± 0.440 % 92.630 ± 0.304 %
|
| 37 |
-
30 4.2686 ± 0.1591 0.01269 ± 0.01350 0.21384 ± 0.01292 11.600 ± 0.432 % 92.745 ± 0.297 %
|
| 38 |
-
31 4.2673 ± 0.1558 0.01106 ± 0.01308 0.20801 ± 0.01252 11.505 ± 0.425 % 92.865 ± 0.290 %
|
| 39 |
-
32 4.1299 ± 0.1470 0.01347 ± 0.01282 0.20576 ± 0.01230 11.498 ± 0.420 % 92.990 ± 0.283 %
|
| 40 |
-
|
| 41 |
-
====== Perplexity statistics ======
|
| 42 |
-
Mean PPL(Q) : 4.129912 ± 0.147024
|
| 43 |
-
Mean PPL(base) : 4.074668 ± 0.139399
|
| 44 |
-
Cor(ln(PPL(Q)), ln(PPL(base))): 93.34%
|
| 45 |
-
Mean ln(PPL(Q)/PPL(base)) : 0.013467 ± 0.012816
|
| 46 |
-
Mean PPL(Q)/PPL(base) : 1.013558 ± 0.012990
|
| 47 |
-
Mean PPL(Q)-PPL(base) : 0.055243 ± 0.052818
|
| 48 |
-
|
| 49 |
-
====== KL divergence statistics ======
|
| 50 |
-
Mean KLD: 0.205759 ± 0.012297
|
| 51 |
-
Maximum KLD: 20.963177
|
| 52 |
-
99.9% KLD: 14.032636
|
| 53 |
-
99.0% KLD: 5.719375
|
| 54 |
-
95.0% KLD: 0.771354
|
| 55 |
-
90.0% KLD: 0.092588
|
| 56 |
-
Median KLD: 0.000773
|
| 57 |
-
10.0% KLD: 0.000002
|
| 58 |
-
5.0% KLD: 0.000000
|
| 59 |
-
1.0% KLD: -0.000000
|
| 60 |
-
0.1% KLD: -0.000001
|
| 61 |
-
Minimum KLD: -0.000013
|
| 62 |
-
|
| 63 |
-
====== Token probability statistics ======
|
| 64 |
-
Mean Δp: -0.023 ± 0.127 %
|
| 65 |
-
Maximum Δp: 99.114%
|
| 66 |
-
99.9% Δp: 96.666%
|
| 67 |
-
99.0% Δp: 42.103%
|
| 68 |
-
95.0% Δp: 4.663%
|
| 69 |
-
90.0% Δp: 1.578%
|
| 70 |
-
75.0% Δp: 0.070%
|
| 71 |
-
Median Δp: -0.000%
|
| 72 |
-
25.0% Δp: -0.066%
|
| 73 |
-
10.0% Δp: -1.639%
|
| 74 |
-
5.0% Δp: -5.594%
|
| 75 |
-
1.0% Δp: -38.947%
|
| 76 |
-
0.1% Δp: -98.266%
|
| 77 |
-
Minimum Δp: -99.912%
|
| 78 |
-
RMS Δp : 11.498 ± 0.420 %
|
| 79 |
-
Same top p: 92.990 ± 0.283 %
|
| 80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|