[config] auto-round W4A16 g128 sym, blocks=['model.language_model.layers', 'mtp.layers'], 16-bit exceptions 97 (expected 97: 48x in_proj_a/b + mtp.fc)
[tensors] source 1199: 407 packed INT4, 792 kept as-is, 0 missing | MTP 15/15 | vision 333 kept | export 19.00 GB
[report] 64 blocks vs L00kUp: mean cos 0.99639 (L00kUp 0.99631), mean psnr 71.0 dB (70.8), iters 46096 (44119), 0 blocks notably worse
   model.language_model.layers.63           cos 0.9888 vs 0.9890  psnr 56.0 vs 56.2
   model.language_model.layers.39           cos 0.9975 vs 0.9975  psnr 68.7 vs 68.7
   model.language_model.layers.0            cos 0.9999 vs 1.0000  psnr 82.0 vs 82.3
   model.language_model.layers.8            cos 0.9999 vs 0.9999  psnr 81.0 vs 80.9
   model.language_model.layers.40           cos 0.9974 vs 0.9974  psnr 68.5 vs 68.4
VERIFY OK