| 0.00.042.837 I common_init_result: fitting params to device memory ... |
| 0.00.042.842 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) |
| 0.00.444.710 W llama_model_loader: direct I/O is enabled, disabling mmap |
| 0.01.459.450 W read_raw_unsafe: Falling back to buffered IO due to Bad address |
| 0.21.377.631 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
| 0.21.428.701 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
| 0.21.707.133 I |
| 0.21.707.313 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 | |
| 0.21.843.964 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1 |
| 0.24.857.762 I kl_divergence: 3.01 seconds per pass - ETA 2.00 minutes |
|
|
| chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p |
| 1 6.1749 ± 0.4909 0.08117 ± 0.01859 0.12763 ± 0.00703 10.392 ± 0.610 % 85.337 ± 1.107 % |
| 2 7.1137 ± 0.3991 0.06516 ± 0.01199 0.10730 ± 0.00416 9.008 ± 0.411 % 86.266 ± 0.761 % |
| 3 7.5580 ± 0.3515 0.07219 ± 0.00958 0.10875 ± 0.00396 8.609 ± 0.316 % 85.989 ± 0.627 % |
| 4 7.8280 ± 0.3239 0.07158 ± 0.00898 0.12059 ± 0.00615 9.140 ± 0.324 % 86.193 ± 0.539 % |
| 5 7.6146 ± 0.2833 0.06405 ± 0.00801 0.11743 ± 0.00502 9.043 ± 0.279 % 86.549 ± 0.477 % |
| 6 6.5667 ± 0.2161 0.06595 ± 0.00747 0.11888 ± 0.00464 9.905 ± 0.299 % 86.804 ± 0.432 % |
| 7 6.0824 ± 0.1827 0.05798 ± 0.00709 0.12603 ± 0.00484 10.284 ± 0.293 % 86.929 ± 0.398 % |
| 8 5.9850 ± 0.1670 0.05416 ± 0.00651 0.12399 ± 0.00430 10.146 ± 0.267 % 86.877 ± 0.373 % |
| 9 6.2982 ± 0.1664 0.05262 ± 0.00612 0.12384 ± 0.00387 9.948 ± 0.246 % 86.445 ± 0.357 % |
| 10 6.4302 ± 0.1624 0.05493 ± 0.00573 0.12014 ± 0.00351 9.781 ± 0.229 % 86.452 ± 0.338 % |
| 11 6.4862 ± 0.1556 0.05482 ± 0.00537 0.11666 ± 0.00323 9.576 ± 0.215 % 86.564 ± 0.322 % |
| 12 6.7484 ± 0.1564 0.05611 ± 0.00505 0.11329 ± 0.00297 9.389 ± 0.203 % 86.470 ± 0.309 % |
| 13 6.7729 ± 0.1504 0.05285 ± 0.00482 0.11177 ± 0.00277 9.295 ± 0.192 % 86.525 ± 0.296 % |
| 14 6.8361 ± 0.1461 0.05335 ± 0.00462 0.11058 ± 0.00260 9.192 ± 0.183 % 86.385 ± 0.287 % |
| 15 6.8710 ± 0.1419 0.05223 ± 0.00443 0.11010 ± 0.00248 9.182 ± 0.175 % 86.373 ± 0.277 % |
| 16 7.0369 ± 0.1408 0.04937 ± 0.00428 0.10941 ± 0.00236 9.094 ± 0.169 % 86.278 ± 0.269 % |
| 17 7.0798 ± 0.1370 0.04848 ± 0.00412 0.10796 ± 0.00224 9.005 ± 0.162 % 86.361 ± 0.260 % |
| 18 7.1674 ± 0.1348 0.04759 ± 0.00399 0.10751 ± 0.00213 8.990 ± 0.157 % 86.201 ± 0.254 % |
| 19 7.1262 ± 0.1310 0.04873 ± 0.00389 0.10651 ± 0.00204 8.894 ± 0.152 % 86.227 ± 0.247 % |
| 20 6.8634 ± 0.1222 0.05049 ± 0.00387 0.11186 ± 0.00203 9.240 ± 0.148 % 86.202 ± 0.241 % |
| 21 6.8908 ± 0.1196 0.05197 ± 0.00379 0.11250 ± 0.00196 9.228 ± 0.144 % 86.175 ± 0.235 % |
| 22 6.9164 ± 0.1174 0.05368 ± 0.00373 0.11404 ± 0.00194 9.255 ± 0.141 % 86.164 ± 0.230 % |
| 23 6.9778 ± 0.1159 0.05511 ± 0.00365 0.11379 ± 0.00188 9.234 ± 0.139 % 86.128 ± 0.225 % |
| 24 6.9773 ± 0.1133 0.05500 ± 0.00356 0.11399 ± 0.00184 9.243 ± 0.137 % 86.038 ± 0.221 % |
| 25 7.0132 ± 0.1116 0.05542 ± 0.00348 0.11365 ± 0.00178 9.211 ± 0.133 % 86.037 ± 0.217 % |
| 26 6.9824 ± 0.1089 0.05549 ± 0.00342 0.11396 ± 0.00176 9.257 ± 0.131 % 86.070 ± 0.212 % |
| 27 7.1525 ± 0.1102 0.05524 ± 0.00337 0.11346 ± 0.00174 9.179 ± 0.129 % 86.072 ± 0.208 % |
| 28 7.2389 ± 0.1098 0.05439 ± 0.00329 0.11219 ± 0.00168 9.107 ± 0.126 % 86.084 ± 0.205 % |
| 29 7.2404 ± 0.1079 0.05545 ± 0.00323 0.11206 ± 0.00164 9.137 ± 0.125 % 86.048 ± 0.201 % |
| 30 7.1812 ± 0.1050 0.05523 ± 0.00318 0.11167 ± 0.00159 9.127 ± 0.122 % 86.022 ± 0.198 % |
| 31 7.0772 ± 0.1015 0.05516 ± 0.00311 0.11097 ± 0.00155 9.104 ± 0.119 % 86.151 ± 0.194 % |
| 32 6.9648 ± 0.0981 0.05342 ± 0.00309 0.11357 ± 0.00169 9.252 ± 0.121 % 86.092 ± 0.191 % |
| 33 6.9013 ± 0.0955 0.05413 ± 0.00306 0.11324 ± 0.00166 9.238 ± 0.119 % 86.087 ± 0.188 % |
| 34 6.8839 ± 0.0937 0.05378 ± 0.00300 0.11241 ± 0.00162 9.209 ± 0.119 % 86.093 ± 0.186 % |
| 35 6.8905 ± 0.0924 0.05293 ± 0.00293 0.11139 ± 0.00158 9.152 ± 0.116 % 86.178 ± 0.182 % |
| 36 6.9081 ± 0.0914 0.05301 ± 0.00290 0.11111 ± 0.00155 9.127 ± 0.114 % 86.190 ± 0.180 % |
| 37 6.8115 ± 0.0887 0.05270 ± 0.00284 0.11014 ± 0.00151 9.096 ± 0.112 % 86.238 ± 0.177 % |
| 38 6.7312 ± 0.0861 0.05138 ± 0.00280 0.10982 ± 0.00148 9.100 ± 0.110 % 86.299 ± 0.174 % |
| 39 6.6496 ± 0.0837 0.05151 ± 0.00276 0.10945 ± 0.00145 9.096 ± 0.108 % 86.312 ± 0.172 % |
| 40 6.5513 ± 0.0811 0.05045 ± 0.00272 0.10876 ± 0.00142 9.101 ± 0.107 % 86.356 ± 0.170 % |
|
|
| ====== Perplexity statistics ====== |
| Mean PPL(Q) : 6.551305 ± 0.081055 |
| Mean PPL(base) : 6.228979 ± 0.075322 |
| Cor(ln(PPL(Q)), ln(PPL(base))): 97.55% |
| Mean ln(PPL(Q)/PPL(base)) : 0.050452 ± 0.002720 |
| Mean PPL(Q)/PPL(base) : 1.051746 ± 0.002861 |
| Mean PPL(Q)-PPL(base) : 0.322326 ± 0.018212 |
|
|
| ====== KL divergence statistics ====== |
| Mean KLD: 0.108756 ± 0.001422 |
| Maximum KLD: 12.410495 |
| 99.9% KLD: 3.419987 |
| 99.0% KLD: 0.930024 |
| 95.0% KLD: 0.366912 |
| 90.0% KLD: 0.235489 |
| Median KLD: 0.050197 |
| 10.0% KLD: 0.000729 |
| 5.0% KLD: 0.000189 |
| 1.0% KLD: -0.000051 |
| 0.1% KLD: -0.000341 |
| Minimum KLD: -0.000842 |
|
|
| ====== Token probability statistics ====== |
| Mean Δp: -0.471 ± 0.045 % |
| Maximum Δp: 96.576% |
| 99.9% Δp: 54.353% |
| 99.0% Δp: 24.394% |
| 95.0% Δp: 11.471% |
| 90.0% Δp: 6.631% |
| 75.0% Δp: 1.298% |
| Median Δp: -0.006% |
| 25.0% Δp: -1.652% |
| 10.0% Δp: -7.598% |
| 5.0% Δp: -13.545% |
| 1.0% Δp: -32.141% |
| 0.1% Δp: -70.328% |
| Minimum Δp: -99.714% |
| RMS Δp : 9.101 ± 0.107 % |
| Same top p: 86.356 ± 0.170 % |
|
|
|
|
