$ C:\quantdiff\llama.cpp\bin\llama-perplexity.exe -m C:\quantdiff\hf-cache\models--bartowski--Qwen_Qwen3.5-4B-GGUF\snapshots\4168f45a16a1290d65a4ec0fa312ae917a4c15d6\Qwen_Qwen3.5-4B-IQ3_M.gguf --kl-divergence-base C:\quantdiff\work\reference\qwen3.5-4b\quantdiff-kld-v1.kld --kl-divergence -c 512 -b 512 -ngl 999 # started 2026-10-21T08:44:45Z 0.00.967.679 W model has unused tensor blk.32.attn_norm.weight (size = 10240 bytes) -- ignoring 0.00.967.694 W model has unused tensor blk.32.post_attention_norm.weight (size = 10240 bytes) -- ignoring 0.00.967.720 W model has unused tensor blk.32.attn_q.weight (size = 22282240 bytes) -- ignoring 0.00.967.729 W model has unused tensor blk.32.attn_k.weight (size = 2785280 bytes) -- ignoring 0.00.967.740 W model has unused tensor blk.32.attn_v.weight (size = 2785280 bytes) -- ignoring 0.00.967.772 W model has unused tensor blk.32.attn_output.weight (size = 11141120 bytes) -- ignoring 0.00.967.783 W model has unused tensor blk.32.attn_q_norm.weight (size = 1024 bytes) -- ignoring 0.00.967.791 W model has unused tensor blk.32.attn_k_norm.weight (size = 1024 bytes) -- ignoring 0.00.967.801 W model has unused tensor blk.32.ffn_gate.weight (size = 25067520 bytes) -- ignoring 0.00.967.810 W model has unused tensor blk.32.ffn_down.weight (size = 25067520 bytes) -- ignoring 0.00.967.821 W model has unused tensor blk.32.ffn_up.weight (size = 25067520 bytes) -- ignoring 0.00.967.831 W model has unused tensor blk.32.nextn.eh_proj.weight (size = 13926400 bytes) -- ignoring 0.00.967.841 W model has unused tensor blk.32.nextn.enorm.weight (size = 10240 bytes) -- ignoring 0.00.967.853 W model has unused tensor blk.32.nextn.hnorm.weight (size = 10240 bytes) -- ignoring 0.00.967.882 W model has unused tensor blk.32.nextn.shared_head_norm.weight (size = 10240 bytes) -- ignoring 0.01.781.493 I cmn init: llama threadpool init, n_threads = 8 0.01.846.200 I 0.01.846.236 I system_info: n_threads = 8 (n_threads_batch = 8) / 16 | CUDA : ARCHS = 750,800,860,890,900,1200,1210 | USE_GRAPHS = 1 | FA_QUANTS = q4_0-q4_0,q8_0-q8_0,f16-f16,bf16-bf16 | BLACKWELL_NATIVE_FP4 = 1 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | 0.01.858.632 I kl_divergence: computing over 100 chunks, n_ctx=512, batch_size=512, n_seq=1 0.02.061.382 I kl_divergence: 0.20 seconds per pass - ETA 0.33 minutes chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p 1 6.6064 ± 1.1357 0.08779 ± 0.03082 0.09344 ± 0.01514 12.395 ± 1.891 % 86.275 ± 2.159 % 2 8.9815 ± 1.1568 0.10583 ± 0.02699 0.10451 ± 0.01224 11.432 ± 1.295 % 85.686 ± 1.552 % 3 7.5335 ± 0.7532 0.09446 ± 0.02115 0.09935 ± 0.00856 10.745 ± 0.957 % 86.144 ± 1.250 % 4 8.8847 ± 0.8052 0.09231 ± 0.01803 0.10348 ± 0.00759 10.253 ± 0.830 % 85.882 ± 1.091 % 5 10.6903 ± 0.8861 0.08269 ± 0.01551 0.09885 ± 0.00621 9.508 ± 0.719 % 84.784 ± 1.006 % 6 10.7593 ± 0.8002 0.07830 ± 0.01368 0.09248 ± 0.00523 8.962 ± 0.638 % 85.359 ± 0.904 % 7 10.7569 ± 0.7303 0.07509 ± 0.01226 0.08799 ± 0.00452 8.574 ± 0.573 % 85.490 ± 0.834 % 8 10.6922 ± 0.6741 0.07325 ± 0.01137 0.08958 ± 0.00424 8.575 ± 0.528 % 86.029 ± 0.768 % 9 11.0038 ± 0.6467 0.07192 ± 0.01039 0.08743 ± 0.00383 8.417 ± 0.493 % 86.405 ± 0.716 % 10 11.7609 ± 0.6618 0.06982 ± 0.00964 0.08578 ± 0.00358 8.134 ± 0.460 % 86.431 ± 0.678 % 11 11.8065 ± 0.6305 0.07252 ± 0.00912 0.08465 ± 0.00328 8.019 ± 0.426 % 85.918 ± 0.657 % 12 12.3398 ± 0.6415 0.07459 ± 0.00867 0.08455 ± 0.00314 7.851 ± 0.401 % 85.882 ± 0.630 % 13 13.1987 ± 0.6669 0.07297 ± 0.00825 0.08351 ± 0.00293 7.729 ± 0.379 % 85.882 ± 0.605 % 14 13.9415 ± 0.6837 0.07591 ± 0.00788 0.08386 ± 0.00275 7.664 ± 0.359 % 85.658 ± 0.587 % 15 14.4057 ± 0.6820 0.07680 ± 0.00755 0.08351 ± 0.00264 7.632 ± 0.351 % 85.516 ± 0.569 % 16 14.3618 ± 0.6566 0.07945 ± 0.00724 0.08235 ± 0.00249 7.575 ± 0.333 % 85.564 ± 0.550 % 17 14.3046 ± 0.6333 0.07799 ± 0.00702 0.08365 ± 0.00241 7.720 ± 0.322 % 85.536 ± 0.534 % 18 14.3720 ± 0.6170 0.07876 ± 0.00696 0.08559 ± 0.00265 7.966 ± 0.335 % 85.359 ± 0.522 % 19 13.1072 ± 0.5433 0.08348 ± 0.00705 0.09084 ± 0.00278 8.802 ± 0.341 % 85.614 ± 0.504 % 20 13.3604 ± 0.5405 0.08320 ± 0.00686 0.09053 ± 0.00268 8.734 ± 0.329 % 85.627 ± 0.491 % 21 12.4668 ± 0.4876 0.08026 ± 0.00660 0.08878 ± 0.00256 8.653 ± 0.317 % 85.920 ± 0.475 % 22 12.4279 ± 0.4752 0.09012 ± 0.00677 0.09465 ± 0.00283 9.209 ± 0.329 % 85.686 ± 0.468 % 23 12.6623 ± 0.4757 0.08832 ± 0.00663 0.09395 ± 0.00273 9.121 ± 0.319 % 85.797 ± 0.456 % 24 11.8661 ± 0.4336 0.09030 ± 0.00647 0.09627 ± 0.00289 9.314 ± 0.307 % 86.029 ± 0.443 % 25 11.5280 ± 0.4108 0.09020 ± 0.00628 0.09520 ± 0.00279 9.285 ± 0.297 % 85.945 ± 0.435 % 26 11.1643 ± 0.3880 0.08793 ± 0.00611 0.09409 ± 0.00268 9.241 ± 0.288 % 86.033 ± 0.426 % 27 11.3094 ± 0.3880 0.08719 ± 0.00604 0.09485 ± 0.00265 9.289 ± 0.285 % 86.042 ± 0.418 % 28 11.2736 ± 0.3796 0.08645 ± 0.00591 0.09485 ± 0.00260 9.273 ± 0.279 % 85.882 ± 0.412 % 29 11.4800 ± 0.3822 0.08806 ± 0.00580 0.09507 ± 0.00257 9.222 ± 0.272 % 85.896 ± 0.405 % 30 11.0171 ± 0.3600 0.09556 ± 0.00599 0.10126 ± 0.00281 10.076 ± 0.291 % 85.908 ± 0.398 % 31 10.9149 ± 0.3489 0.09542 ± 0.00587 0.10060 ± 0.00273 10.002 ± 0.284 % 85.832 ± 0.392 % 32 10.9971 ± 0.3485 0.09436 ± 0.00620 0.10862 ± 0.00364 10.100 ± 0.287 % 85.625 ± 0.388 % 33 10.8969 ± 0.3393 0.09328 ± 0.00609 0.10841 ± 0.00354 10.071 ± 0.280 % 85.668 ± 0.382 % 34 10.8552 ± 0.3318 0.09167 ± 0.00597 0.10784 ± 0.00345 10.045 ± 0.274 % 85.652 ± 0.377 % 35 10.5302 ± 0.3157 0.09114 ± 0.00585 0.10702 ± 0.00335 10.059 ± 0.267 % 85.669 ± 0.371 % 36 10.2505 ± 0.3017 0.09435 ± 0.00581 0.10920 ± 0.00333 10.310 ± 0.265 % 85.577 ± 0.367 % 37 9.8584 ± 0.2846 0.09282 ± 0.00569 0.10792 ± 0.00324 10.288 ± 0.259 % 85.787 ± 0.360 % 38 9.7364 ± 0.2764 0.09197 ± 0.00559 0.10777 ± 0.00317 10.325 ± 0.253 % 85.851 ± 0.354 % 39 9.7684 ± 0.2742 0.09499 ± 0.00581 0.10986 ± 0.00341 10.327 ± 0.251 % 85.842 ± 0.350 % 40 9.6113 ± 0.2652 0.09604 ± 0.00573 0.10974 ± 0.00334 10.279 ± 0.246 % 85.843 ± 0.345 % 41 9.5953 ± 0.2608 0.09682 ± 0.00565 0.10920 ± 0.00326 10.240 ± 0.242 % 85.882 ± 0.341 % 42 9.4218 ± 0.2522 0.09536 ± 0.00555 0.10815 ± 0.00319 10.196 ± 0.238 % 85.938 ± 0.336 % 43 9.2984 ± 0.2467 0.09424 ± 0.00547 0.10741 ± 0.00312 10.156 ± 0.234 % 86.001 ± 0.331 % 44 9.1939 ± 0.2402 0.09377 ± 0.00538 0.10688 ± 0.00306 10.147 ± 0.230 % 86.105 ± 0.327 % 45 9.0564 ± 0.2340 0.09257 ± 0.00528 0.10550 ± 0.00299 10.084 ± 0.226 % 86.275 ± 0.321 % 46 9.1345 ± 0.2334 0.09267 ± 0.00521 0.10511 ± 0.00293 10.036 ± 0.223 % 86.240 ± 0.318 % 47 9.1063 ± 0.2298 0.09257 ± 0.00514 0.10454 ± 0.00287 9.996 ± 0.220 % 86.224 ± 0.315 % 48 8.9587 ± 0.2226 0.09273 ± 0.00506 0.10367 ± 0.00282 9.954 ± 0.216 % 86.266 ± 0.311 % 49 9.0002 ± 0.2224 0.09358 ± 0.00507 0.10357 ± 0.00277 9.923 ± 0.214 % 86.267 ± 0.308 % 50 9.0053 ± 0.2200 0.09371 ± 0.00500 0.10337 ± 0.00272 9.894 ± 0.211 % 86.267 ± 0.305 % 51 8.9289 ± 0.2152 0.09157 ± 0.00496 0.10421 ± 0.00275 9.848 ± 0.208 % 86.282 ± 0.302 % 52 9.0273 ± 0.2159 0.09025 ± 0.00493 0.10504 ± 0.00278 9.851 ± 0.206 % 86.237 ± 0.299 % 53 9.0650 ± 0.2143 0.09046 ± 0.00488 0.10490 ± 0.00273 9.827 ± 0.203 % 86.119 ± 0.297 % 54 9.0885 ± 0.2130 0.08986 ± 0.00482 0.10460 ± 0.00269 9.796 ± 0.200 % 86.166 ± 0.294 % 55 9.1664 ± 0.2137 0.09048 ± 0.00477 0.10429 ± 0.00264 9.802 ± 0.198 % 86.160 ± 0.292 % 56 9.1563 ± 0.2110 0.08922 ± 0.00471 0.10372 ± 0.00260 9.758 ± 0.195 % 86.148 ± 0.289 % 57 9.2957 ± 0.2127 0.08947 ± 0.00467 0.10380 ± 0.00256 9.762 ± 0.193 % 86.082 ± 0.287 % 58 9.2821 ± 0.2103 0.08937 ± 0.00461 0.10338 ± 0.00251 9.733 ± 0.191 % 86.112 ± 0.284 % 59 9.2482 ± 0.2077 0.08853 ± 0.00456 0.10312 ± 0.00248 9.719 ± 0.188 % 86.162 ± 0.282 % 60 9.1898 ± 0.2052 0.08915 ± 0.00454 0.10304 ± 0.00244 9.754 ± 0.187 % 86.176 ± 0.279 % 61 9.2480 ± 0.2054 0.08923 ± 0.00449 0.10278 ± 0.00241 9.734 ± 0.186 % 86.178 ± 0.277 % 62 9.2093 ± 0.2024 0.08837 ± 0.00444 0.10217 ± 0.00237 9.689 ± 0.184 % 86.186 ± 0.274 % 63 9.2231 ± 0.2011 0.08799 ± 0.00439 0.10204 ± 0.00234 9.691 ± 0.183 % 86.212 ± 0.272 % 64 9.2683 ± 0.2003 0.08759 ± 0.00435 0.10172 ± 0.00231 9.659 ± 0.181 % 86.170 ± 0.270 % 65 9.0918 ± 0.1945 0.08779 ± 0.00435 0.10267 ± 0.00230 9.774 ± 0.179 % 86.250 ± 0.267 % 66 9.1338 ± 0.1941 0.08755 ± 0.00430 0.10245 ± 0.00227 9.753 ± 0.177 % 86.209 ± 0.266 % 67 9.1267 ± 0.1920 0.08740 ± 0.00425 0.10207 ± 0.00224 9.711 ± 0.176 % 86.181 ± 0.264 % 68 9.1254 ± 0.1904 0.08654 ± 0.00421 0.10162 ± 0.00220 9.679 ± 0.174 % 86.205 ± 0.262 % 69 9.1681 ± 0.1899 0.08677 ± 0.00417 0.10132 ± 0.00217 9.658 ± 0.172 % 86.235 ± 0.260 % 70 9.1094 ± 0.1873 0.08617 ± 0.00414 0.10123 ± 0.00216 9.640 ± 0.170 % 86.325 ± 0.257 % 71 9.0109 ± 0.1840 0.08556 ± 0.00410 0.10070 ± 0.00213 9.623 ± 0.169 % 86.418 ± 0.255 % 72 9.1322 ± 0.1856 0.08465 ± 0.00406 0.10061 ± 0.00211 9.604 ± 0.168 % 86.411 ± 0.253 % 73 9.1453 ± 0.1846 0.08438 ± 0.00402 0.10020 ± 0.00208 9.571 ± 0.166 % 86.414 ± 0.251 % 74 9.1900 ± 0.1843 0.08445 ± 0.00398 0.09983 ± 0.00205 9.543 ± 0.165 % 86.444 ± 0.249 % 75 9.1266 ± 0.1816 0.08386 ± 0.00394 0.09996 ± 0.00204 9.538 ± 0.163 % 86.447 ± 0.248 % 76 9.1031 ± 0.1800 0.08481 ± 0.00394 0.10017 ± 0.00202 9.557 ± 0.162 % 86.450 ± 0.246 % 77 9.0424 ± 0.1771 0.08371 ± 0.00390 0.10054 ± 0.00207 9.547 ± 0.160 % 86.453 ± 0.244 % 78 9.0716 ± 0.1767 0.08383 ± 0.00387 0.10028 ± 0.00204 9.534 ± 0.159 % 86.420 ± 0.243 % 79 9.0332 ± 0.1746 0.08329 ± 0.00384 0.09994 ± 0.00202 9.526 ± 0.157 % 86.463 ± 0.241 % 80 9.0310 ± 0.1735 0.08348 ± 0.00382 0.10088 ± 0.00201 9.542 ± 0.156 % 86.412 ± 0.240 % 81 9.0988 ± 0.1740 0.08351 ± 0.00380 0.10085 ± 0.00199 9.519 ± 0.154 % 86.396 ± 0.239 % 82 9.0918 ± 0.1728 0.08339 ± 0.00378 0.10126 ± 0.00201 9.489 ± 0.153 % 86.418 ± 0.237 % 83 9.1860 ± 0.1737 0.08306 ± 0.00374 0.10106 ± 0.00199 9.462 ± 0.152 % 86.374 ± 0.236 % 84 9.1994 ± 0.1727 0.08288 ± 0.00371 0.10074 ± 0.00197 9.442 ± 0.151 % 86.396 ± 0.234 % 85 9.1156 ± 0.1699 0.08285 ± 0.00370 0.10082 ± 0.00195 9.434 ± 0.149 % 86.436 ± 0.233 % 86 9.2133 ± 0.1709 0.08185 ± 0.00367 0.10059 ± 0.00193 9.407 ± 0.148 % 86.434 ± 0.231 % 87 9.1853 ± 0.1691 0.08211 ± 0.00367 0.10054 ± 0.00191 9.393 ± 0.147 % 86.378 ± 0.230 % 88 9.2040 ± 0.1684 0.08199 ± 0.00364 0.10022 ± 0.00189 9.376 ± 0.146 % 86.395 ± 0.229 % 89 9.1957 ± 0.1671 0.08169 ± 0.00361 0.09977 ± 0.00187 9.352 ± 0.145 % 86.393 ± 0.228 % 90 9.1810 ± 0.1658 0.08094 ± 0.00358 0.09989 ± 0.00186 9.344 ± 0.143 % 86.401 ± 0.226 % 91 9.2939 ± 0.1676 0.08104 ± 0.00357 0.10023 ± 0.00184 9.320 ± 0.142 % 86.387 ± 0.225 % 92 9.2826 ± 0.1664 0.08048 ± 0.00355 0.10022 ± 0.00183 9.308 ± 0.141 % 86.360 ± 0.224 % 93 9.3345 ± 0.1666 0.07978 ± 0.00352 0.09996 ± 0.00181 9.296 ± 0.140 % 86.338 ± 0.223 % 94 9.3687 ± 0.1663 0.07941 ± 0.00350 0.09974 ± 0.00179 9.274 ± 0.139 % 86.354 ± 0.222 % 95 9.3845 ± 0.1657 0.07924 ± 0.00348 0.09969 ± 0.00177 9.258 ± 0.138 % 86.398 ± 0.220 % 96 9.4033 ± 0.1653 0.07962 ± 0.00348 0.09987 ± 0.00176 9.257 ± 0.137 % 86.389 ± 0.219 % 97 9.4287 ± 0.1648 0.07959 ± 0.00346 0.10015 ± 0.00176 9.232 ± 0.136 % 86.355 ± 0.218 % 98 9.5165 ± 0.1658 0.07945 ± 0.00345 0.10026 ± 0.00175 9.215 ± 0.135 % 86.323 ± 0.217 % 99 9.5443 ± 0.1653 0.07920 ± 0.00342 0.09998 ± 0.00173 9.199 ± 0.134 % 86.342 ± 0.216 % 100 9.5979 ± 0.1656 0.07960 ± 0.00340 0.10023 ± 0.00172 9.198 ± 0.134 % 86.314 ± 0.215 % ====== Perplexity statistics ====== Mean PPL(Q) : 9.597882 ± 0.165591 Mean PPL(base) : 8.863479 ± 0.153159 Cor(ln(PPL(Q)), ln(PPL(base))): 98.06% Mean ln(PPL(Q)/PPL(base)) : 0.079603 ± 0.003405 Mean PPL(Q)/PPL(base) : 1.082857 ± 0.003687 Mean PPL(Q)-PPL(base) : 0.734403 ± 0.033772 ====== KL divergence statistics ====== Mean KLD: 0.100231 ± 0.001724 Maximum KLD: 12.282552 99.9% KLD: 3.255949 99.0% KLD: 0.929547 95.0% KLD: 0.297636 90.0% KLD: 0.192190 Median KLD: 0.051890 10.0% KLD: 0.003559 5.0% KLD: 0.001207 1.0% KLD: 0.000176 0.1% KLD: 0.000029 Minimum KLD: 0.000000 ====== Token probability statistics ====== Mean Δp: -2.488 ± 0.055 % Maximum Δp: 97.081% 99.9% Δp: 40.198% 99.0% Δp: 17.220% 95.0% Δp: 6.466% 90.0% Δp: 3.025% 75.0% Δp: 0.201% Median Δp: -0.337% 25.0% Δp: -3.924% 10.0% Δp: -10.795% 5.0% Δp: -16.605% 1.0% Δp: -35.948% 0.1% Δp: -74.922% Minimum Δp: -99.762% RMS Δp : 9.198 ± 0.134 % Same top p: 86.314 ± 0.215 % # exit 0 after 19s