| 0.00.037.317 I common_init_result: fitting params to device memory ... |
| 0.00.037.321 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on) |
| 0.00.390.383 W llama_model_loader: direct I/O is enabled, disabling mmap |
| 0.21.870.791 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
| 0.21.936.318 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
| 0.22.143.467 I |
| 0.22.143.580 I system_info: n_threads = 16 (n_threads_batch = 16) / 32 | ROCm : NO_VMM = 1 | PEER_MAX_BATCH_SIZE = 128 | FA_ALL_QUANTS = 1 | 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 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | |
| 0.22.258.917 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1 |
| 0.24.621.838 I kl_divergence: 2.36 seconds per pass - ETA 1.57 minutes |
|
|
| chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p |
| 1 5.6299 ± 0.4284 -0.01123 ± 0.01762 0.10095 ± 0.00634 9.176 ± 0.605 % 88.270 ± 1.007 % |
| 2 6.6587 ± 0.3601 -0.00093 ± 0.01117 0.08359 ± 0.00360 7.978 ± 0.395 % 88.025 ± 0.718 % |
| 3 7.1385 ± 0.3226 0.01508 ± 0.00875 0.08242 ± 0.00314 7.696 ± 0.308 % 87.911 ± 0.589 % |
| 4 7.4207 ± 0.2990 0.01815 ± 0.00760 0.08463 ± 0.00438 7.777 ± 0.308 % 88.465 ± 0.499 % |
| 5 7.2795 ± 0.2642 0.01905 ± 0.00665 0.08239 ± 0.00359 7.768 ± 0.267 % 88.426 ± 0.447 % |
| 6 6.2745 ± 0.2014 0.02044 ± 0.00615 0.08353 ± 0.00339 8.576 ± 0.294 % 88.987 ± 0.400 % |
| 7 5.8589 ± 0.1719 0.02053 ± 0.00613 0.09196 ± 0.00373 9.221 ± 0.305 % 88.926 ± 0.371 % |
| 8 5.7865 ± 0.1577 0.02044 ± 0.00558 0.09026 ± 0.00333 9.081 ± 0.275 % 88.734 ± 0.350 % |
| 9 6.0902 ± 0.1574 0.01904 ± 0.00522 0.09017 ± 0.00300 8.885 ± 0.253 % 88.530 ± 0.332 % |
| 10 6.2036 ± 0.1532 0.01905 ± 0.00484 0.08664 ± 0.00271 8.652 ± 0.235 % 88.592 ± 0.314 % |
| 11 6.2550 ± 0.1467 0.01852 ± 0.00454 0.08389 ± 0.00248 8.471 ± 0.220 % 88.661 ± 0.299 % |
| 12 6.5002 ± 0.1472 0.01864 ± 0.00426 0.08096 ± 0.00228 8.271 ± 0.208 % 88.702 ± 0.286 % |
| 13 6.5547 ± 0.1424 0.02010 ± 0.00407 0.07947 ± 0.00212 8.179 ± 0.196 % 88.668 ± 0.275 % |
| 14 6.6139 ± 0.1383 0.02031 ± 0.00387 0.07794 ± 0.00199 8.054 ± 0.186 % 88.773 ± 0.264 % |
| 15 6.6603 ± 0.1347 0.02109 ± 0.00372 0.07724 ± 0.00186 7.982 ± 0.177 % 88.739 ± 0.255 % |
| 16 6.8133 ± 0.1336 0.01708 ± 0.00360 0.07665 ± 0.00178 7.924 ± 0.170 % 88.612 ± 0.248 % |
| 17 6.8557 ± 0.1300 0.01631 ± 0.00347 0.07576 ± 0.00169 7.851 ± 0.164 % 88.615 ± 0.241 % |
| 18 6.9493 ± 0.1282 0.01668 ± 0.00337 0.07548 ± 0.00160 7.811 ± 0.158 % 88.541 ± 0.235 % |
| 19 6.9014 ± 0.1244 0.01668 ± 0.00325 0.07445 ± 0.00154 7.725 ± 0.153 % 88.645 ± 0.228 % |
| 20 6.6526 ± 0.1161 0.01930 ± 0.00323 0.07848 ± 0.00152 7.994 ± 0.146 % 88.514 ± 0.223 % |
| 21 6.6703 ± 0.1135 0.01945 ± 0.00316 0.07876 ± 0.00146 7.982 ± 0.141 % 88.423 ± 0.218 % |
| 22 6.6837 ± 0.1111 0.01946 ± 0.00308 0.07952 ± 0.00143 8.019 ± 0.138 % 88.376 ± 0.214 % |
| 23 6.7415 ± 0.1096 0.02066 ± 0.00300 0.07952 ± 0.00138 7.991 ± 0.134 % 88.278 ± 0.210 % |
| 24 6.7373 ± 0.1070 0.02000 ± 0.00295 0.07956 ± 0.00138 7.986 ± 0.133 % 88.270 ± 0.205 % |
| 25 6.7719 ± 0.1054 0.02041 ± 0.00289 0.07925 ± 0.00134 7.953 ± 0.130 % 88.258 ± 0.201 % |
| 26 6.7420 ± 0.1028 0.02046 ± 0.00283 0.07951 ± 0.00131 7.991 ± 0.129 % 88.274 ± 0.197 % |
| 27 6.9113 ± 0.1041 0.02093 ± 0.00283 0.07927 ± 0.00136 7.933 ± 0.126 % 88.248 ± 0.194 % |
| 28 7.0022 ± 0.1039 0.02114 ± 0.00276 0.07854 ± 0.00132 7.873 ± 0.123 % 88.256 ± 0.190 % |
| 29 6.9981 ± 0.1020 0.02141 ± 0.00271 0.07833 ± 0.00128 7.882 ± 0.121 % 88.250 ± 0.187 % |
| 30 6.9536 ± 0.0995 0.02303 ± 0.00267 0.07848 ± 0.00125 7.892 ± 0.119 % 88.244 ± 0.184 % |
| 31 6.8548 ± 0.0962 0.02323 ± 0.00261 0.07786 ± 0.00122 7.868 ± 0.117 % 88.320 ± 0.180 % |
| 32 6.7489 ± 0.0931 0.02193 ± 0.00261 0.08043 ± 0.00142 8.006 ± 0.117 % 88.273 ± 0.178 % |
| 33 6.6865 ± 0.0906 0.02252 ± 0.00256 0.07990 ± 0.00139 7.989 ± 0.115 % 88.299 ± 0.175 % |
| 34 6.6723 ± 0.0889 0.02256 ± 0.00251 0.07907 ± 0.00135 7.928 ± 0.113 % 88.365 ± 0.172 % |
| 35 6.6827 ± 0.0878 0.02231 ± 0.00246 0.07849 ± 0.00131 7.908 ± 0.111 % 88.404 ± 0.169 % |
| 36 6.6954 ± 0.0868 0.02172 ± 0.00242 0.07826 ± 0.00128 7.892 ± 0.110 % 88.414 ± 0.167 % |
| 37 6.6026 ± 0.0841 0.02157 ± 0.00238 0.07763 ± 0.00125 7.875 ± 0.108 % 88.407 ± 0.165 % |
| 38 6.5309 ± 0.0818 0.02118 ± 0.00235 0.07760 ± 0.00123 7.895 ± 0.106 % 88.432 ± 0.162 % |
| 39 6.4522 ± 0.0795 0.02138 ± 0.00232 0.07746 ± 0.00121 7.905 ± 0.105 % 88.423 ± 0.160 % |
| 40 6.3601 ± 0.0770 0.02082 ± 0.00229 0.07695 ± 0.00118 7.909 ± 0.104 % 88.463 ± 0.158 % |
|
|
| ====== Perplexity statistics ====== |
| Mean PPL(Q) : 6.360050 ± 0.077045 |
| Mean PPL(base) : 6.228979 ± 0.075322 |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.21% |
| Mean ln(PPL(Q)/PPL(base)) : 0.020824 ± 0.002287 |
| Mean PPL(Q)/PPL(base) : 1.021042 ± 0.002336 |
| Mean PPL(Q)-PPL(base) : 0.131071 ± 0.014499 |
|
|
| ====== KL divergence statistics ====== |
| Mean KLD: 0.076947 ± 0.001181 |
| Maximum KLD: 16.293211 |
| 99.9% KLD: 2.499913 |
| 99.0% KLD: 0.685225 |
| 95.0% KLD: 0.259541 |
| 90.0% KLD: 0.162843 |
| Median KLD: 0.034241 |
| 10.0% KLD: 0.000500 |
| 5.0% KLD: 0.000134 |
| 1.0% KLD: -0.000030 |
| 0.1% KLD: -0.000274 |
| Minimum KLD: -0.000654 |
|
|
| ====== Token probability statistics ====== |
| Mean Δp: -0.505 ± 0.039 % |
| Maximum Δp: 98.233% |
| 99.9% Δp: 50.439% |
| 99.0% Δp: 21.309% |
| 95.0% Δp: 9.386% |
| 90.0% Δp: 5.261% |
| 75.0% Δp: 0.918% |
| Median Δp: -0.011% |
| 25.0% Δp: -1.620% |
| 10.0% Δp: -6.708% |
| 5.0% Δp: -11.584% |
| 1.0% Δp: -26.637% |
| 0.1% Δp: -64.007% |
| Minimum Δp: -99.856% |
| RMS Δp : 7.909 ± 0.104 % |
| Same top p: 88.463 ± 0.158 % |
|
|
|
|
