| 0.00.035.061 I common_init_result: fitting params to device memory ... |
| 0.00.035.064 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) |
| 0.00.364.958 W llama_model_loader: direct I/O is enabled, disabling mmap |
| 0.01.365.200 W read_raw_unsafe: Falling back to buffered IO due to Bad address |
| 0.20.724.587 W llama_context: n_ctx_seq (2048) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
| 0.20.776.655 W common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
| 0.20.958.449 I |
| 0.20.958.535 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 | |
| 0.21.072.467 I kl_divergence: computing over 40 chunks, n_ctx=2048, batch_size=2048, n_seq=1 |
| 0.23.478.059 I kl_divergence: 2.41 seconds per pass - ETA 1.60 minutes |
|
|
| chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p |
| 1 5.5867 ± 0.4228 -0.01894 ± 0.01754 0.10256 ± 0.00648 9.538 ± 0.610 % 88.368 ± 1.003 % |
| 2 6.6450 ± 0.3596 -0.00300 ± 0.01113 0.08488 ± 0.00372 8.059 ± 0.404 % 88.270 ± 0.712 % |
| 3 7.0875 ± 0.3197 0.00792 ± 0.00873 0.08299 ± 0.00332 7.716 ± 0.310 % 88.172 ± 0.583 % |
| 4 7.3455 ± 0.2950 0.00796 ± 0.00746 0.08559 ± 0.00479 7.862 ± 0.310 % 88.759 ± 0.494 % |
| 5 7.2194 ± 0.2615 0.01075 ± 0.00655 0.08256 ± 0.00391 7.727 ± 0.267 % 88.524 ± 0.446 % |
| 6 6.2395 ± 0.1998 0.01485 ± 0.00603 0.08240 ± 0.00357 8.402 ± 0.286 % 88.970 ± 0.400 % |
| 7 5.8467 ± 0.1714 0.01845 ± 0.00617 0.09250 ± 0.00438 9.020 ± 0.301 % 88.884 ± 0.371 % |
| 8 5.7766 ± 0.1573 0.01872 ± 0.00560 0.09000 ± 0.00387 8.841 ± 0.272 % 88.869 ± 0.348 % |
| 9 6.0753 ± 0.1568 0.01659 ± 0.00524 0.08968 ± 0.00347 8.651 ± 0.250 % 88.617 ± 0.331 % |
| 10 6.1938 ± 0.1528 0.01748 ± 0.00486 0.08628 ± 0.00314 8.443 ± 0.232 % 88.778 ± 0.312 % |
| 11 6.2507 ± 0.1465 0.01783 ± 0.00456 0.08369 ± 0.00287 8.284 ± 0.217 % 88.812 ± 0.297 % |
| 12 6.4984 ± 0.1471 0.01836 ± 0.00428 0.08087 ± 0.00264 8.100 ± 0.205 % 88.840 ± 0.284 % |
| 13 6.5527 ± 0.1424 0.01980 ± 0.00408 0.07955 ± 0.00245 8.022 ± 0.193 % 88.646 ± 0.275 % |
| 14 6.6129 ± 0.1383 0.02016 ± 0.00389 0.07798 ± 0.00228 7.902 ± 0.183 % 88.752 ± 0.264 % |
| 15 6.6619 ± 0.1348 0.02133 ± 0.00373 0.07730 ± 0.00214 7.842 ± 0.174 % 88.693 ± 0.256 % |
| 16 6.8209 ± 0.1338 0.01819 ± 0.00361 0.07652 ± 0.00203 7.773 ± 0.167 % 88.618 ± 0.248 % |
| 17 6.8676 ± 0.1304 0.01805 ± 0.00348 0.07551 ± 0.00192 7.696 ± 0.162 % 88.609 ± 0.241 % |
| 18 6.9618 ± 0.1286 0.01849 ± 0.00338 0.07510 ± 0.00182 7.653 ± 0.155 % 88.552 ± 0.235 % |
| 19 6.9170 ± 0.1249 0.01894 ± 0.00325 0.07404 ± 0.00175 7.561 ± 0.151 % 88.661 ± 0.227 % |
| 20 6.6681 ± 0.1166 0.02161 ± 0.00323 0.07796 ± 0.00171 7.836 ± 0.145 % 88.548 ± 0.223 % |
| 21 6.6815 ± 0.1138 0.02112 ± 0.00316 0.07831 ± 0.00165 7.820 ± 0.139 % 88.428 ± 0.218 % |
| 22 6.6992 ± 0.1116 0.02177 ± 0.00313 0.07958 ± 0.00169 7.862 ± 0.139 % 88.394 ± 0.214 % |
| 23 6.7561 ± 0.1101 0.02283 ± 0.00306 0.07988 ± 0.00164 7.864 ± 0.136 % 88.270 ± 0.210 % |
| 24 6.7528 ± 0.1074 0.02230 ± 0.00301 0.07975 ± 0.00161 7.848 ± 0.134 % 88.262 ± 0.205 % |
| 25 6.7834 ± 0.1058 0.02210 ± 0.00294 0.07935 ± 0.00155 7.808 ± 0.131 % 88.223 ± 0.202 % |
| 26 6.7551 ± 0.1032 0.02241 ± 0.00288 0.07966 ± 0.00152 7.846 ± 0.129 % 88.270 ± 0.197 % |
| 27 6.9239 ± 0.1045 0.02275 ± 0.00288 0.07942 ± 0.00156 7.792 ± 0.127 % 88.244 ± 0.194 % |
| 28 7.0172 ± 0.1043 0.02329 ± 0.00282 0.07891 ± 0.00153 7.749 ± 0.124 % 88.259 ± 0.190 % |
| 29 7.0127 ± 0.1024 0.02349 ± 0.00277 0.07872 ± 0.00148 7.761 ± 0.122 % 88.263 ± 0.187 % |
| 30 6.9664 ± 0.0999 0.02487 ± 0.00272 0.07871 ± 0.00144 7.761 ± 0.119 % 88.263 ± 0.184 % |
| 31 6.8666 ± 0.0966 0.02496 ± 0.00265 0.07794 ± 0.00140 7.729 ± 0.117 % 88.377 ± 0.180 % |
| 32 6.7575 ± 0.0934 0.02320 ± 0.00266 0.08099 ± 0.00163 7.910 ± 0.120 % 88.313 ± 0.178 % |
| 33 6.6952 ± 0.0909 0.02382 ± 0.00260 0.08035 ± 0.00159 7.884 ± 0.118 % 88.371 ± 0.174 % |
| 34 6.6814 ± 0.0892 0.02392 ± 0.00254 0.07951 ± 0.00154 7.831 ± 0.115 % 88.416 ± 0.172 % |
| 35 6.6908 ± 0.0880 0.02351 ± 0.00249 0.07890 ± 0.00150 7.807 ± 0.113 % 88.465 ± 0.169 % |
| 36 6.7037 ± 0.0870 0.02296 ± 0.00245 0.07859 ± 0.00146 7.786 ± 0.111 % 88.498 ± 0.166 % |
| 37 6.6096 ± 0.0844 0.02263 ± 0.00241 0.07791 ± 0.00143 7.765 ± 0.109 % 88.492 ± 0.164 % |
| 38 6.5383 ± 0.0821 0.02230 ± 0.00238 0.07778 ± 0.00140 7.790 ± 0.108 % 88.514 ± 0.162 % |
| 39 6.4567 ± 0.0797 0.02207 ± 0.00234 0.07744 ± 0.00136 7.798 ± 0.106 % 88.508 ± 0.160 % |
| 40 6.3632 ± 0.0772 0.02132 ± 0.00231 0.07684 ± 0.00133 7.783 ± 0.104 % 88.556 ± 0.157 % |
|
|
| ====== Perplexity statistics ====== |
| Mean PPL(Q) : 6.363210 ± 0.077202 |
| Mean PPL(base) : 6.228979 ± 0.075322 |
| Cor(ln(PPL(Q)), ln(PPL(base))): 98.19% |
| Mean ln(PPL(Q)/PPL(base)) : 0.021321 ± 0.002307 |
| Mean PPL(Q)/PPL(base) : 1.021549 ± 0.002357 |
| Mean PPL(Q)-PPL(base) : 0.134231 ± 0.014643 |
|
|
| ====== KL divergence statistics ====== |
| Mean KLD: 0.076844 ± 0.001332 |
| Maximum KLD: 16.142012 |
| 99.9% KLD: 2.816024 |
| 99.0% KLD: 0.669352 |
| 95.0% KLD: 0.255460 |
| 90.0% KLD: 0.161866 |
| Median KLD: 0.033844 |
| 10.0% KLD: 0.000495 |
| 5.0% KLD: 0.000131 |
| 1.0% KLD: -0.000030 |
| 0.1% KLD: -0.000287 |
| Minimum KLD: -0.000637 |
|
|
| ====== Token probability statistics ====== |
| Mean Δp: -0.458 ± 0.038 % |
| Maximum Δp: 99.702% |
| 99.9% Δp: 52.654% |
| 99.0% Δp: 20.917% |
| 95.0% Δp: 9.373% |
| 90.0% Δp: 5.248% |
| 75.0% Δp: 0.940% |
| Median Δp: -0.010% |
| 25.0% Δp: -1.540% |
| 10.0% Δp: -6.629% |
| 5.0% Δp: -11.288% |
| 1.0% Δp: -25.973% |
| 0.1% Δp: -60.301% |
| Minimum Δp: -99.838% |
| RMS Δp : 7.783 ± 0.104 % |
| Same top p: 88.556 ± 0.157 % |
|
|
|
|
